1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG. This pass is where
12 // algebraic simplification happens.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #define DEBUG_TYPE "instcombine"
37 #include "llvm/Transforms/Scalar.h"
38 #include "llvm/IntrinsicInst.h"
39 #include "llvm/Pass.h"
40 #include "llvm/DerivedTypes.h"
41 #include "llvm/GlobalVariable.h"
42 #include "llvm/Analysis/ConstantFolding.h"
43 #include "llvm/Analysis/ValueTracking.h"
44 #include "llvm/Target/TargetData.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Support/CallSite.h"
48 #include "llvm/Support/ConstantRange.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/GetElementPtrTypeIterator.h"
51 #include "llvm/Support/InstVisitor.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Support/PatternMatch.h"
54 #include "llvm/Support/Compiler.h"
55 #include "llvm/ADT/DenseMap.h"
56 #include "llvm/ADT/SmallVector.h"
57 #include "llvm/ADT/SmallPtrSet.h"
58 #include "llvm/ADT/Statistic.h"
59 #include "llvm/ADT/STLExtras.h"
64 using namespace llvm::PatternMatch;
66 STATISTIC(NumCombined , "Number of insts combined");
67 STATISTIC(NumConstProp, "Number of constant folds");
68 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
69 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
70 STATISTIC(NumSunkInst , "Number of instructions sunk");
73 class VISIBILITY_HIDDEN InstCombiner
74 : public FunctionPass,
75 public InstVisitor<InstCombiner, Instruction*> {
76 // Worklist of all of the instructions that need to be simplified.
77 SmallVector<Instruction*, 256> Worklist;
78 DenseMap<Instruction*, unsigned> WorklistMap;
80 bool MustPreserveLCSSA;
82 static char ID; // Pass identification, replacement for typeid
83 InstCombiner() : FunctionPass(&ID) {}
85 /// AddToWorkList - Add the specified instruction to the worklist if it
86 /// isn't already in it.
87 void AddToWorkList(Instruction *I) {
88 if (WorklistMap.insert(std::make_pair(I, Worklist.size())).second)
89 Worklist.push_back(I);
92 // RemoveFromWorkList - remove I from the worklist if it exists.
93 void RemoveFromWorkList(Instruction *I) {
94 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
95 if (It == WorklistMap.end()) return; // Not in worklist.
97 // Don't bother moving everything down, just null out the slot.
98 Worklist[It->second] = 0;
100 WorklistMap.erase(It);
103 Instruction *RemoveOneFromWorkList() {
104 Instruction *I = Worklist.back();
106 WorklistMap.erase(I);
111 /// AddUsersToWorkList - When an instruction is simplified, add all users of
112 /// the instruction to the work lists because they might get more simplified
115 void AddUsersToWorkList(Value &I) {
116 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
118 AddToWorkList(cast<Instruction>(*UI));
121 /// AddUsesToWorkList - When an instruction is simplified, add operands to
122 /// the work lists because they might get more simplified now.
124 void AddUsesToWorkList(Instruction &I) {
125 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
126 if (Instruction *Op = dyn_cast<Instruction>(*i))
130 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
131 /// dead. Add all of its operands to the worklist, turning them into
132 /// undef's to reduce the number of uses of those instructions.
134 /// Return the specified operand before it is turned into an undef.
136 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
137 Value *R = I.getOperand(op);
139 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
140 if (Instruction *Op = dyn_cast<Instruction>(*i)) {
142 // Set the operand to undef to drop the use.
143 *i = UndefValue::get(Op->getType());
150 virtual bool runOnFunction(Function &F);
152 bool DoOneIteration(Function &F, unsigned ItNum);
154 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
155 AU.addRequired<TargetData>();
156 AU.addPreservedID(LCSSAID);
157 AU.setPreservesCFG();
160 TargetData &getTargetData() const { return *TD; }
162 // Visitation implementation - Implement instruction combining for different
163 // instruction types. The semantics are as follows:
165 // null - No change was made
166 // I - Change was made, I is still valid, I may be dead though
167 // otherwise - Change was made, replace I with returned instruction
169 Instruction *visitAdd(BinaryOperator &I);
170 Instruction *visitSub(BinaryOperator &I);
171 Instruction *visitMul(BinaryOperator &I);
172 Instruction *visitURem(BinaryOperator &I);
173 Instruction *visitSRem(BinaryOperator &I);
174 Instruction *visitFRem(BinaryOperator &I);
175 bool SimplifyDivRemOfSelect(BinaryOperator &I);
176 Instruction *commonRemTransforms(BinaryOperator &I);
177 Instruction *commonIRemTransforms(BinaryOperator &I);
178 Instruction *commonDivTransforms(BinaryOperator &I);
179 Instruction *commonIDivTransforms(BinaryOperator &I);
180 Instruction *visitUDiv(BinaryOperator &I);
181 Instruction *visitSDiv(BinaryOperator &I);
182 Instruction *visitFDiv(BinaryOperator &I);
183 Instruction *FoldAndOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
184 Instruction *visitAnd(BinaryOperator &I);
185 Instruction *FoldOrOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
186 Instruction *FoldOrWithConstants(BinaryOperator &I, Value *Op,
187 Value *A, Value *B, Value *C);
188 Instruction *visitOr (BinaryOperator &I);
189 Instruction *visitXor(BinaryOperator &I);
190 Instruction *visitShl(BinaryOperator &I);
191 Instruction *visitAShr(BinaryOperator &I);
192 Instruction *visitLShr(BinaryOperator &I);
193 Instruction *commonShiftTransforms(BinaryOperator &I);
194 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
196 Instruction *visitFCmpInst(FCmpInst &I);
197 Instruction *visitICmpInst(ICmpInst &I);
198 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
199 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
202 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
203 ConstantInt *DivRHS);
205 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
206 ICmpInst::Predicate Cond, Instruction &I);
207 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
209 Instruction *commonCastTransforms(CastInst &CI);
210 Instruction *commonIntCastTransforms(CastInst &CI);
211 Instruction *commonPointerCastTransforms(CastInst &CI);
212 Instruction *visitTrunc(TruncInst &CI);
213 Instruction *visitZExt(ZExtInst &CI);
214 Instruction *visitSExt(SExtInst &CI);
215 Instruction *visitFPTrunc(FPTruncInst &CI);
216 Instruction *visitFPExt(CastInst &CI);
217 Instruction *visitFPToUI(FPToUIInst &FI);
218 Instruction *visitFPToSI(FPToSIInst &FI);
219 Instruction *visitUIToFP(CastInst &CI);
220 Instruction *visitSIToFP(CastInst &CI);
221 Instruction *visitPtrToInt(CastInst &CI);
222 Instruction *visitIntToPtr(IntToPtrInst &CI);
223 Instruction *visitBitCast(BitCastInst &CI);
224 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
226 Instruction *visitSelectInst(SelectInst &SI);
227 Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
228 Instruction *visitCallInst(CallInst &CI);
229 Instruction *visitInvokeInst(InvokeInst &II);
230 Instruction *visitPHINode(PHINode &PN);
231 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
232 Instruction *visitAllocationInst(AllocationInst &AI);
233 Instruction *visitFreeInst(FreeInst &FI);
234 Instruction *visitLoadInst(LoadInst &LI);
235 Instruction *visitStoreInst(StoreInst &SI);
236 Instruction *visitBranchInst(BranchInst &BI);
237 Instruction *visitSwitchInst(SwitchInst &SI);
238 Instruction *visitInsertElementInst(InsertElementInst &IE);
239 Instruction *visitExtractElementInst(ExtractElementInst &EI);
240 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
241 Instruction *visitExtractValueInst(ExtractValueInst &EV);
243 // visitInstruction - Specify what to return for unhandled instructions...
244 Instruction *visitInstruction(Instruction &I) { return 0; }
247 Instruction *visitCallSite(CallSite CS);
248 bool transformConstExprCastCall(CallSite CS);
249 Instruction *transformCallThroughTrampoline(CallSite CS);
250 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
251 bool DoXform = true);
252 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
253 DbgDeclareInst *hasOneUsePlusDeclare(Value *V);
257 // InsertNewInstBefore - insert an instruction New before instruction Old
258 // in the program. Add the new instruction to the worklist.
260 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
261 assert(New && New->getParent() == 0 &&
262 "New instruction already inserted into a basic block!");
263 BasicBlock *BB = Old.getParent();
264 BB->getInstList().insert(&Old, New); // Insert inst
269 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
270 /// This also adds the cast to the worklist. Finally, this returns the
272 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
274 if (V->getType() == Ty) return V;
276 if (Constant *CV = dyn_cast<Constant>(V))
277 return ConstantExpr::getCast(opc, CV, Ty);
279 Instruction *C = CastInst::Create(opc, V, Ty, V->getName(), &Pos);
284 Value *InsertBitCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
285 return InsertCastBefore(Instruction::BitCast, V, Ty, Pos);
289 // ReplaceInstUsesWith - This method is to be used when an instruction is
290 // found to be dead, replacable with another preexisting expression. Here
291 // we add all uses of I to the worklist, replace all uses of I with the new
292 // value, then return I, so that the inst combiner will know that I was
295 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
296 AddUsersToWorkList(I); // Add all modified instrs to worklist
298 I.replaceAllUsesWith(V);
301 // If we are replacing the instruction with itself, this must be in a
302 // segment of unreachable code, so just clobber the instruction.
303 I.replaceAllUsesWith(UndefValue::get(I.getType()));
308 // EraseInstFromFunction - When dealing with an instruction that has side
309 // effects or produces a void value, we can't rely on DCE to delete the
310 // instruction. Instead, visit methods should return the value returned by
312 Instruction *EraseInstFromFunction(Instruction &I) {
313 assert(I.use_empty() && "Cannot erase instruction that is used!");
314 AddUsesToWorkList(I);
315 RemoveFromWorkList(&I);
317 return 0; // Don't do anything with FI
320 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt &KnownZero,
321 APInt &KnownOne, unsigned Depth = 0) const {
322 return llvm::ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
325 bool MaskedValueIsZero(Value *V, const APInt &Mask,
326 unsigned Depth = 0) const {
327 return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
329 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
330 return llvm::ComputeNumSignBits(Op, TD, Depth);
335 /// SimplifyCommutative - This performs a few simplifications for
336 /// commutative operators.
337 bool SimplifyCommutative(BinaryOperator &I);
339 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
340 /// most-complex to least-complex order.
341 bool SimplifyCompare(CmpInst &I);
343 /// SimplifyDemandedUseBits - Attempts to replace V with a simpler value
344 /// based on the demanded bits.
345 Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
346 APInt& KnownZero, APInt& KnownOne,
348 bool SimplifyDemandedBits(Use &U, APInt DemandedMask,
349 APInt& KnownZero, APInt& KnownOne,
352 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
353 /// SimplifyDemandedBits knows about. See if the instruction has any
354 /// properties that allow us to simplify its operands.
355 bool SimplifyDemandedInstructionBits(Instruction &Inst);
357 Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
358 APInt& UndefElts, unsigned Depth = 0);
360 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
361 // PHI node as operand #0, see if we can fold the instruction into the PHI
362 // (which is only possible if all operands to the PHI are constants).
363 Instruction *FoldOpIntoPhi(Instruction &I);
365 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
366 // operator and they all are only used by the PHI, PHI together their
367 // inputs, and do the operation once, to the result of the PHI.
368 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
369 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
370 Instruction *FoldPHIArgGEPIntoPHI(PHINode &PN);
373 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
374 ConstantInt *AndRHS, BinaryOperator &TheAnd);
376 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
377 bool isSub, Instruction &I);
378 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
379 bool isSigned, bool Inside, Instruction &IB);
380 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
381 Instruction *MatchBSwap(BinaryOperator &I);
382 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
383 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
384 Instruction *SimplifyMemSet(MemSetInst *MI);
387 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
389 bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
390 unsigned CastOpc, int &NumCastsRemoved);
391 unsigned GetOrEnforceKnownAlignment(Value *V,
392 unsigned PrefAlign = 0);
397 char InstCombiner::ID = 0;
398 static RegisterPass<InstCombiner>
399 X("instcombine", "Combine redundant instructions");
401 // getComplexity: Assign a complexity or rank value to LLVM Values...
402 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
403 static unsigned getComplexity(Value *V) {
404 if (isa<Instruction>(V)) {
405 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
409 if (isa<Argument>(V)) return 3;
410 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
413 // isOnlyUse - Return true if this instruction will be deleted if we stop using
415 static bool isOnlyUse(Value *V) {
416 return V->hasOneUse() || isa<Constant>(V);
419 // getPromotedType - Return the specified type promoted as it would be to pass
420 // though a va_arg area...
421 static const Type *getPromotedType(const Type *Ty) {
422 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
423 if (ITy->getBitWidth() < 32)
424 return Type::Int32Ty;
429 /// getBitCastOperand - If the specified operand is a CastInst, a constant
430 /// expression bitcast, or a GetElementPtrInst with all zero indices, return the
431 /// operand value, otherwise return null.
432 static Value *getBitCastOperand(Value *V) {
433 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
435 return I->getOperand(0);
436 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
437 // GetElementPtrInst?
438 if (GEP->hasAllZeroIndices())
439 return GEP->getOperand(0);
440 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
441 if (CE->getOpcode() == Instruction::BitCast)
442 // BitCast ConstantExp?
443 return CE->getOperand(0);
444 else if (CE->getOpcode() == Instruction::GetElementPtr) {
445 // GetElementPtr ConstantExp?
446 for (User::op_iterator I = CE->op_begin() + 1, E = CE->op_end();
448 ConstantInt *CI = dyn_cast<ConstantInt>(I);
449 if (!CI || !CI->isZero())
450 // Any non-zero indices? Not cast-like.
453 // All-zero indices? This is just like casting.
454 return CE->getOperand(0);
460 /// This function is a wrapper around CastInst::isEliminableCastPair. It
461 /// simply extracts arguments and returns what that function returns.
462 static Instruction::CastOps
463 isEliminableCastPair(
464 const CastInst *CI, ///< The first cast instruction
465 unsigned opcode, ///< The opcode of the second cast instruction
466 const Type *DstTy, ///< The target type for the second cast instruction
467 TargetData *TD ///< The target data for pointer size
470 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
471 const Type *MidTy = CI->getType(); // B from above
473 // Get the opcodes of the two Cast instructions
474 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
475 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
477 return Instruction::CastOps(
478 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
479 DstTy, TD->getIntPtrType()));
482 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
483 /// in any code being generated. It does not require codegen if V is simple
484 /// enough or if the cast can be folded into other casts.
485 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
486 const Type *Ty, TargetData *TD) {
487 if (V->getType() == Ty || isa<Constant>(V)) return false;
489 // If this is another cast that can be eliminated, it isn't codegen either.
490 if (const CastInst *CI = dyn_cast<CastInst>(V))
491 if (isEliminableCastPair(CI, opcode, Ty, TD))
496 // SimplifyCommutative - This performs a few simplifications for commutative
499 // 1. Order operands such that they are listed from right (least complex) to
500 // left (most complex). This puts constants before unary operators before
503 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
504 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
506 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
507 bool Changed = false;
508 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
509 Changed = !I.swapOperands();
511 if (!I.isAssociative()) return Changed;
512 Instruction::BinaryOps Opcode = I.getOpcode();
513 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
514 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
515 if (isa<Constant>(I.getOperand(1))) {
516 Constant *Folded = ConstantExpr::get(I.getOpcode(),
517 cast<Constant>(I.getOperand(1)),
518 cast<Constant>(Op->getOperand(1)));
519 I.setOperand(0, Op->getOperand(0));
520 I.setOperand(1, Folded);
522 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
523 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
524 isOnlyUse(Op) && isOnlyUse(Op1)) {
525 Constant *C1 = cast<Constant>(Op->getOperand(1));
526 Constant *C2 = cast<Constant>(Op1->getOperand(1));
528 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
529 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
530 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
534 I.setOperand(0, New);
535 I.setOperand(1, Folded);
542 /// SimplifyCompare - For a CmpInst this function just orders the operands
543 /// so that theyare listed from right (least complex) to left (most complex).
544 /// This puts constants before unary operators before binary operators.
545 bool InstCombiner::SimplifyCompare(CmpInst &I) {
546 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
549 // Compare instructions are not associative so there's nothing else we can do.
553 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
554 // if the LHS is a constant zero (which is the 'negate' form).
556 static inline Value *dyn_castNegVal(Value *V) {
557 if (BinaryOperator::isNeg(V))
558 return BinaryOperator::getNegArgument(V);
560 // Constants can be considered to be negated values if they can be folded.
561 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
562 return ConstantExpr::getNeg(C);
564 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
565 if (C->getType()->getElementType()->isInteger())
566 return ConstantExpr::getNeg(C);
571 static inline Value *dyn_castNotVal(Value *V) {
572 if (BinaryOperator::isNot(V))
573 return BinaryOperator::getNotArgument(V);
575 // Constants can be considered to be not'ed values...
576 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
577 return ConstantInt::get(~C->getValue());
581 // dyn_castFoldableMul - If this value is a multiply that can be folded into
582 // other computations (because it has a constant operand), return the
583 // non-constant operand of the multiply, and set CST to point to the multiplier.
584 // Otherwise, return null.
586 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
587 if (V->hasOneUse() && V->getType()->isInteger())
588 if (Instruction *I = dyn_cast<Instruction>(V)) {
589 if (I->getOpcode() == Instruction::Mul)
590 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
591 return I->getOperand(0);
592 if (I->getOpcode() == Instruction::Shl)
593 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
594 // The multiplier is really 1 << CST.
595 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
596 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
597 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
598 return I->getOperand(0);
604 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
605 /// expression, return it.
606 static User *dyn_castGetElementPtr(Value *V) {
607 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
608 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
609 if (CE->getOpcode() == Instruction::GetElementPtr)
610 return cast<User>(V);
614 /// getOpcode - If this is an Instruction or a ConstantExpr, return the
615 /// opcode value. Otherwise return UserOp1.
616 static unsigned getOpcode(const Value *V) {
617 if (const Instruction *I = dyn_cast<Instruction>(V))
618 return I->getOpcode();
619 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
620 return CE->getOpcode();
621 // Use UserOp1 to mean there's no opcode.
622 return Instruction::UserOp1;
625 /// AddOne - Add one to a ConstantInt
626 static ConstantInt *AddOne(ConstantInt *C) {
627 APInt Val(C->getValue());
628 return ConstantInt::get(++Val);
630 /// SubOne - Subtract one from a ConstantInt
631 static ConstantInt *SubOne(ConstantInt *C) {
632 APInt Val(C->getValue());
633 return ConstantInt::get(--Val);
635 /// Add - Add two ConstantInts together
636 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
637 return ConstantInt::get(C1->getValue() + C2->getValue());
639 /// And - Bitwise AND two ConstantInts together
640 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
641 return ConstantInt::get(C1->getValue() & C2->getValue());
643 /// Subtract - Subtract one ConstantInt from another
644 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
645 return ConstantInt::get(C1->getValue() - C2->getValue());
647 /// Multiply - Multiply two ConstantInts together
648 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
649 return ConstantInt::get(C1->getValue() * C2->getValue());
651 /// MultiplyOverflows - True if the multiply can not be expressed in an int
653 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
654 uint32_t W = C1->getBitWidth();
655 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
664 APInt MulExt = LHSExt * RHSExt;
667 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
668 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
669 return MulExt.slt(Min) || MulExt.sgt(Max);
671 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
675 /// ShrinkDemandedConstant - Check to see if the specified operand of the
676 /// specified instruction is a constant integer. If so, check to see if there
677 /// are any bits set in the constant that are not demanded. If so, shrink the
678 /// constant and return true.
679 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
681 assert(I && "No instruction?");
682 assert(OpNo < I->getNumOperands() && "Operand index too large");
684 // If the operand is not a constant integer, nothing to do.
685 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
686 if (!OpC) return false;
688 // If there are no bits set that aren't demanded, nothing to do.
689 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
690 if ((~Demanded & OpC->getValue()) == 0)
693 // This instruction is producing bits that are not demanded. Shrink the RHS.
694 Demanded &= OpC->getValue();
695 I->setOperand(OpNo, ConstantInt::get(Demanded));
699 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
700 // set of known zero and one bits, compute the maximum and minimum values that
701 // could have the specified known zero and known one bits, returning them in
703 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
704 const APInt& KnownZero,
705 const APInt& KnownOne,
706 APInt& Min, APInt& Max) {
707 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
708 assert(KnownZero.getBitWidth() == BitWidth &&
709 KnownOne.getBitWidth() == BitWidth &&
710 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
711 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
712 APInt UnknownBits = ~(KnownZero|KnownOne);
714 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
715 // bit if it is unknown.
717 Max = KnownOne|UnknownBits;
719 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
721 Max.clear(BitWidth-1);
725 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
726 // a set of known zero and one bits, compute the maximum and minimum values that
727 // could have the specified known zero and known one bits, returning them in
729 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
730 const APInt &KnownZero,
731 const APInt &KnownOne,
732 APInt &Min, APInt &Max) {
733 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
734 assert(KnownZero.getBitWidth() == BitWidth &&
735 KnownOne.getBitWidth() == BitWidth &&
736 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
737 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
738 APInt UnknownBits = ~(KnownZero|KnownOne);
740 // The minimum value is when the unknown bits are all zeros.
742 // The maximum value is when the unknown bits are all ones.
743 Max = KnownOne|UnknownBits;
746 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
747 /// SimplifyDemandedBits knows about. See if the instruction has any
748 /// properties that allow us to simplify its operands.
749 bool InstCombiner::SimplifyDemandedInstructionBits(Instruction &Inst) {
750 unsigned BitWidth = cast<IntegerType>(Inst.getType())->getBitWidth();
751 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
752 APInt DemandedMask(APInt::getAllOnesValue(BitWidth));
754 Value *V = SimplifyDemandedUseBits(&Inst, DemandedMask,
755 KnownZero, KnownOne, 0);
756 if (V == 0) return false;
757 if (V == &Inst) return true;
758 ReplaceInstUsesWith(Inst, V);
762 /// SimplifyDemandedBits - This form of SimplifyDemandedBits simplifies the
763 /// specified instruction operand if possible, updating it in place. It returns
764 /// true if it made any change and false otherwise.
765 bool InstCombiner::SimplifyDemandedBits(Use &U, APInt DemandedMask,
766 APInt &KnownZero, APInt &KnownOne,
768 Value *NewVal = SimplifyDemandedUseBits(U.get(), DemandedMask,
769 KnownZero, KnownOne, Depth);
770 if (NewVal == 0) return false;
776 /// SimplifyDemandedUseBits - This function attempts to replace V with a simpler
777 /// value based on the demanded bits. When this function is called, it is known
778 /// that only the bits set in DemandedMask of the result of V are ever used
779 /// downstream. Consequently, depending on the mask and V, it may be possible
780 /// to replace V with a constant or one of its operands. In such cases, this
781 /// function does the replacement and returns true. In all other cases, it
782 /// returns false after analyzing the expression and setting KnownOne and known
783 /// to be one in the expression. KnownZero contains all the bits that are known
784 /// to be zero in the expression. These are provided to potentially allow the
785 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
786 /// the expression. KnownOne and KnownZero always follow the invariant that
787 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
788 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
789 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
790 /// and KnownOne must all be the same.
792 /// This returns null if it did not change anything and it permits no
793 /// simplification. This returns V itself if it did some simplification of V's
794 /// operands based on the information about what bits are demanded. This returns
795 /// some other non-null value if it found out that V is equal to another value
796 /// in the context where the specified bits are demanded, but not for all users.
797 Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
798 APInt &KnownZero, APInt &KnownOne,
800 assert(V != 0 && "Null pointer of Value???");
801 assert(Depth <= 6 && "Limit Search Depth");
802 uint32_t BitWidth = DemandedMask.getBitWidth();
803 const IntegerType *VTy = cast<IntegerType>(V->getType());
804 assert(VTy->getBitWidth() == BitWidth &&
805 KnownZero.getBitWidth() == BitWidth &&
806 KnownOne.getBitWidth() == BitWidth &&
807 "Value *V, DemandedMask, KnownZero and KnownOne \
808 must have same BitWidth");
809 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
810 // We know all of the bits for a constant!
811 KnownOne = CI->getValue() & DemandedMask;
812 KnownZero = ~KnownOne & DemandedMask;
818 if (DemandedMask == 0) { // Not demanding any bits from V.
819 if (isa<UndefValue>(V))
821 return UndefValue::get(VTy);
824 if (Depth == 6) // Limit search depth.
827 Instruction *I = dyn_cast<Instruction>(V);
828 if (!I) return 0; // Only analyze instructions.
830 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
831 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
833 // If there are multiple uses of this value and we aren't at the root, then
834 // we can't do any simplifications of the operands, because DemandedMask
835 // only reflects the bits demanded by *one* of the users.
836 if (Depth != 0 && !I->hasOneUse()) {
837 // Despite the fact that we can't simplify this instruction in all User's
838 // context, we can at least compute the knownzero/knownone bits, and we can
839 // do simplifications that apply to *just* the one user if we know that
840 // this instruction has a simpler value in that context.
841 if (I->getOpcode() == Instruction::And) {
842 // If either the LHS or the RHS are Zero, the result is zero.
843 ComputeMaskedBits(I->getOperand(1), DemandedMask,
844 RHSKnownZero, RHSKnownOne, Depth+1);
845 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
846 LHSKnownZero, LHSKnownOne, Depth+1);
848 // If all of the demanded bits are known 1 on one side, return the other.
849 // These bits cannot contribute to the result of the 'and' in this
851 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
852 (DemandedMask & ~LHSKnownZero))
853 return I->getOperand(0);
854 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
855 (DemandedMask & ~RHSKnownZero))
856 return I->getOperand(1);
858 // If all of the demanded bits in the inputs are known zeros, return zero.
859 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
860 return Constant::getNullValue(VTy);
862 } else if (I->getOpcode() == Instruction::Or) {
863 // We can simplify (X|Y) -> X or Y in the user's context if we know that
864 // only bits from X or Y are demanded.
866 // If either the LHS or the RHS are One, the result is One.
867 ComputeMaskedBits(I->getOperand(1), DemandedMask,
868 RHSKnownZero, RHSKnownOne, Depth+1);
869 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
870 LHSKnownZero, LHSKnownOne, Depth+1);
872 // If all of the demanded bits are known zero on one side, return the
873 // other. These bits cannot contribute to the result of the 'or' in this
875 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
876 (DemandedMask & ~LHSKnownOne))
877 return I->getOperand(0);
878 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
879 (DemandedMask & ~RHSKnownOne))
880 return I->getOperand(1);
882 // If all of the potentially set bits on one side are known to be set on
883 // the other side, just use the 'other' side.
884 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
885 (DemandedMask & (~RHSKnownZero)))
886 return I->getOperand(0);
887 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
888 (DemandedMask & (~LHSKnownZero)))
889 return I->getOperand(1);
892 // Compute the KnownZero/KnownOne bits to simplify things downstream.
893 ComputeMaskedBits(I, DemandedMask, KnownZero, KnownOne, Depth);
897 // If this is the root being simplified, allow it to have multiple uses,
898 // just set the DemandedMask to all bits so that we can try to simplify the
899 // operands. This allows visitTruncInst (for example) to simplify the
900 // operand of a trunc without duplicating all the logic below.
901 if (Depth == 0 && !V->hasOneUse())
902 DemandedMask = APInt::getAllOnesValue(BitWidth);
904 switch (I->getOpcode()) {
906 ComputeMaskedBits(I, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
908 case Instruction::And:
909 // If either the LHS or the RHS are Zero, the result is zero.
910 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
911 RHSKnownZero, RHSKnownOne, Depth+1) ||
912 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownZero,
913 LHSKnownZero, LHSKnownOne, Depth+1))
915 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
916 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
918 // If all of the demanded bits are known 1 on one side, return the other.
919 // These bits cannot contribute to the result of the 'and'.
920 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
921 (DemandedMask & ~LHSKnownZero))
922 return I->getOperand(0);
923 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
924 (DemandedMask & ~RHSKnownZero))
925 return I->getOperand(1);
927 // If all of the demanded bits in the inputs are known zeros, return zero.
928 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
929 return Constant::getNullValue(VTy);
931 // If the RHS is a constant, see if we can simplify it.
932 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
935 // Output known-1 bits are only known if set in both the LHS & RHS.
936 RHSKnownOne &= LHSKnownOne;
937 // Output known-0 are known to be clear if zero in either the LHS | RHS.
938 RHSKnownZero |= LHSKnownZero;
940 case Instruction::Or:
941 // If either the LHS or the RHS are One, the result is One.
942 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
943 RHSKnownZero, RHSKnownOne, Depth+1) ||
944 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownOne,
945 LHSKnownZero, LHSKnownOne, Depth+1))
947 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
948 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
950 // If all of the demanded bits are known zero on one side, return the other.
951 // These bits cannot contribute to the result of the 'or'.
952 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
953 (DemandedMask & ~LHSKnownOne))
954 return I->getOperand(0);
955 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
956 (DemandedMask & ~RHSKnownOne))
957 return I->getOperand(1);
959 // If all of the potentially set bits on one side are known to be set on
960 // the other side, just use the 'other' side.
961 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
962 (DemandedMask & (~RHSKnownZero)))
963 return I->getOperand(0);
964 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
965 (DemandedMask & (~LHSKnownZero)))
966 return I->getOperand(1);
968 // If the RHS is a constant, see if we can simplify it.
969 if (ShrinkDemandedConstant(I, 1, DemandedMask))
972 // Output known-0 bits are only known if clear in both the LHS & RHS.
973 RHSKnownZero &= LHSKnownZero;
974 // Output known-1 are known to be set if set in either the LHS | RHS.
975 RHSKnownOne |= LHSKnownOne;
977 case Instruction::Xor: {
978 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
979 RHSKnownZero, RHSKnownOne, Depth+1) ||
980 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
981 LHSKnownZero, LHSKnownOne, Depth+1))
983 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
984 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
986 // If all of the demanded bits are known zero on one side, return the other.
987 // These bits cannot contribute to the result of the 'xor'.
988 if ((DemandedMask & RHSKnownZero) == DemandedMask)
989 return I->getOperand(0);
990 if ((DemandedMask & LHSKnownZero) == DemandedMask)
991 return I->getOperand(1);
993 // Output known-0 bits are known if clear or set in both the LHS & RHS.
994 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
995 (RHSKnownOne & LHSKnownOne);
996 // Output known-1 are known to be set if set in only one of the LHS, RHS.
997 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
998 (RHSKnownOne & LHSKnownZero);
1000 // If all of the demanded bits are known to be zero on one side or the
1001 // other, turn this into an *inclusive* or.
1002 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1003 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1005 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1007 return InsertNewInstBefore(Or, *I);
1010 // If all of the demanded bits on one side are known, and all of the set
1011 // bits on that side are also known to be set on the other side, turn this
1012 // into an AND, as we know the bits will be cleared.
1013 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1014 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1016 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1017 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
1019 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
1020 return InsertNewInstBefore(And, *I);
1024 // If the RHS is a constant, see if we can simplify it.
1025 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1026 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1029 RHSKnownZero = KnownZeroOut;
1030 RHSKnownOne = KnownOneOut;
1033 case Instruction::Select:
1034 if (SimplifyDemandedBits(I->getOperandUse(2), DemandedMask,
1035 RHSKnownZero, RHSKnownOne, Depth+1) ||
1036 SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1037 LHSKnownZero, LHSKnownOne, Depth+1))
1039 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1040 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1042 // If the operands are constants, see if we can simplify them.
1043 if (ShrinkDemandedConstant(I, 1, DemandedMask) ||
1044 ShrinkDemandedConstant(I, 2, DemandedMask))
1047 // Only known if known in both the LHS and RHS.
1048 RHSKnownOne &= LHSKnownOne;
1049 RHSKnownZero &= LHSKnownZero;
1051 case Instruction::Trunc: {
1052 unsigned truncBf = I->getOperand(0)->getType()->getPrimitiveSizeInBits();
1053 DemandedMask.zext(truncBf);
1054 RHSKnownZero.zext(truncBf);
1055 RHSKnownOne.zext(truncBf);
1056 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1057 RHSKnownZero, RHSKnownOne, Depth+1))
1059 DemandedMask.trunc(BitWidth);
1060 RHSKnownZero.trunc(BitWidth);
1061 RHSKnownOne.trunc(BitWidth);
1062 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1065 case Instruction::BitCast:
1066 if (!I->getOperand(0)->getType()->isInteger())
1067 return false; // vector->int or fp->int?
1068 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1069 RHSKnownZero, RHSKnownOne, Depth+1))
1071 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1073 case Instruction::ZExt: {
1074 // Compute the bits in the result that are not present in the input.
1075 unsigned SrcBitWidth =I->getOperand(0)->getType()->getPrimitiveSizeInBits();
1077 DemandedMask.trunc(SrcBitWidth);
1078 RHSKnownZero.trunc(SrcBitWidth);
1079 RHSKnownOne.trunc(SrcBitWidth);
1080 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1081 RHSKnownZero, RHSKnownOne, Depth+1))
1083 DemandedMask.zext(BitWidth);
1084 RHSKnownZero.zext(BitWidth);
1085 RHSKnownOne.zext(BitWidth);
1086 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1087 // The top bits are known to be zero.
1088 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1091 case Instruction::SExt: {
1092 // Compute the bits in the result that are not present in the input.
1093 unsigned SrcBitWidth =I->getOperand(0)->getType()->getPrimitiveSizeInBits();
1095 APInt InputDemandedBits = DemandedMask &
1096 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1098 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1099 // If any of the sign extended bits are demanded, we know that the sign
1101 if ((NewBits & DemandedMask) != 0)
1102 InputDemandedBits.set(SrcBitWidth-1);
1104 InputDemandedBits.trunc(SrcBitWidth);
1105 RHSKnownZero.trunc(SrcBitWidth);
1106 RHSKnownOne.trunc(SrcBitWidth);
1107 if (SimplifyDemandedBits(I->getOperandUse(0), InputDemandedBits,
1108 RHSKnownZero, RHSKnownOne, Depth+1))
1110 InputDemandedBits.zext(BitWidth);
1111 RHSKnownZero.zext(BitWidth);
1112 RHSKnownOne.zext(BitWidth);
1113 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1115 // If the sign bit of the input is known set or clear, then we know the
1116 // top bits of the result.
1118 // If the input sign bit is known zero, or if the NewBits are not demanded
1119 // convert this into a zero extension.
1120 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits) {
1121 // Convert to ZExt cast
1122 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName());
1123 return InsertNewInstBefore(NewCast, *I);
1124 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1125 RHSKnownOne |= NewBits;
1129 case Instruction::Add: {
1130 // Figure out what the input bits are. If the top bits of the and result
1131 // are not demanded, then the add doesn't demand them from its input
1133 unsigned NLZ = DemandedMask.countLeadingZeros();
1135 // If there is a constant on the RHS, there are a variety of xformations
1137 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1138 // If null, this should be simplified elsewhere. Some of the xforms here
1139 // won't work if the RHS is zero.
1143 // If the top bit of the output is demanded, demand everything from the
1144 // input. Otherwise, we demand all the input bits except NLZ top bits.
1145 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1147 // Find information about known zero/one bits in the input.
1148 if (SimplifyDemandedBits(I->getOperandUse(0), InDemandedBits,
1149 LHSKnownZero, LHSKnownOne, Depth+1))
1152 // If the RHS of the add has bits set that can't affect the input, reduce
1154 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1157 // Avoid excess work.
1158 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1161 // Turn it into OR if input bits are zero.
1162 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1164 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1166 return InsertNewInstBefore(Or, *I);
1169 // We can say something about the output known-zero and known-one bits,
1170 // depending on potential carries from the input constant and the
1171 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1172 // bits set and the RHS constant is 0x01001, then we know we have a known
1173 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1175 // To compute this, we first compute the potential carry bits. These are
1176 // the bits which may be modified. I'm not aware of a better way to do
1178 const APInt &RHSVal = RHS->getValue();
1179 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1181 // Now that we know which bits have carries, compute the known-1/0 sets.
1183 // Bits are known one if they are known zero in one operand and one in the
1184 // other, and there is no input carry.
1185 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1186 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1188 // Bits are known zero if they are known zero in both operands and there
1189 // is no input carry.
1190 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1192 // If the high-bits of this ADD are not demanded, then it does not demand
1193 // the high bits of its LHS or RHS.
1194 if (DemandedMask[BitWidth-1] == 0) {
1195 // Right fill the mask of bits for this ADD to demand the most
1196 // significant bit and all those below it.
1197 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1198 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1199 LHSKnownZero, LHSKnownOne, Depth+1) ||
1200 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1201 LHSKnownZero, LHSKnownOne, Depth+1))
1207 case Instruction::Sub:
1208 // If the high-bits of this SUB are not demanded, then it does not demand
1209 // the high bits of its LHS or RHS.
1210 if (DemandedMask[BitWidth-1] == 0) {
1211 // Right fill the mask of bits for this SUB to demand the most
1212 // significant bit and all those below it.
1213 uint32_t NLZ = DemandedMask.countLeadingZeros();
1214 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1215 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1216 LHSKnownZero, LHSKnownOne, Depth+1) ||
1217 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1218 LHSKnownZero, LHSKnownOne, Depth+1))
1221 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1222 // the known zeros and ones.
1223 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1225 case Instruction::Shl:
1226 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1227 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1228 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1229 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1230 RHSKnownZero, RHSKnownOne, Depth+1))
1232 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1233 RHSKnownZero <<= ShiftAmt;
1234 RHSKnownOne <<= ShiftAmt;
1235 // low bits known zero.
1237 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1240 case Instruction::LShr:
1241 // For a logical shift right
1242 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1243 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1245 // Unsigned shift right.
1246 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1247 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1248 RHSKnownZero, RHSKnownOne, Depth+1))
1250 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1251 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1252 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1254 // Compute the new bits that are at the top now.
1255 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1256 RHSKnownZero |= HighBits; // high bits known zero.
1260 case Instruction::AShr:
1261 // If this is an arithmetic shift right and only the low-bit is set, we can
1262 // always convert this into a logical shr, even if the shift amount is
1263 // variable. The low bit of the shift cannot be an input sign bit unless
1264 // the shift amount is >= the size of the datatype, which is undefined.
1265 if (DemandedMask == 1) {
1266 // Perform the logical shift right.
1267 Instruction *NewVal = BinaryOperator::CreateLShr(
1268 I->getOperand(0), I->getOperand(1), I->getName());
1269 return InsertNewInstBefore(NewVal, *I);
1272 // If the sign bit is the only bit demanded by this ashr, then there is no
1273 // need to do it, the shift doesn't change the high bit.
1274 if (DemandedMask.isSignBit())
1275 return I->getOperand(0);
1277 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1278 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1280 // Signed shift right.
1281 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1282 // If any of the "high bits" are demanded, we should set the sign bit as
1284 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1285 DemandedMaskIn.set(BitWidth-1);
1286 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1287 RHSKnownZero, RHSKnownOne, Depth+1))
1289 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1290 // Compute the new bits that are at the top now.
1291 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1292 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1293 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1295 // Handle the sign bits.
1296 APInt SignBit(APInt::getSignBit(BitWidth));
1297 // Adjust to where it is now in the mask.
1298 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1300 // If the input sign bit is known to be zero, or if none of the top bits
1301 // are demanded, turn this into an unsigned shift right.
1302 if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] ||
1303 (HighBits & ~DemandedMask) == HighBits) {
1304 // Perform the logical shift right.
1305 Instruction *NewVal = BinaryOperator::CreateLShr(
1306 I->getOperand(0), SA, I->getName());
1307 return InsertNewInstBefore(NewVal, *I);
1308 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1309 RHSKnownOne |= HighBits;
1313 case Instruction::SRem:
1314 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1315 APInt RA = Rem->getValue().abs();
1316 if (RA.isPowerOf2()) {
1317 if (DemandedMask.ule(RA)) // srem won't affect demanded bits
1318 return I->getOperand(0);
1320 APInt LowBits = RA - 1;
1321 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1322 if (SimplifyDemandedBits(I->getOperandUse(0), Mask2,
1323 LHSKnownZero, LHSKnownOne, Depth+1))
1326 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1327 LHSKnownZero |= ~LowBits;
1329 KnownZero |= LHSKnownZero & DemandedMask;
1331 assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
1335 case Instruction::URem: {
1336 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1337 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1338 if (SimplifyDemandedBits(I->getOperandUse(0), AllOnes,
1339 KnownZero2, KnownOne2, Depth+1) ||
1340 SimplifyDemandedBits(I->getOperandUse(1), AllOnes,
1341 KnownZero2, KnownOne2, Depth+1))
1344 unsigned Leaders = KnownZero2.countLeadingOnes();
1345 Leaders = std::max(Leaders,
1346 KnownZero2.countLeadingOnes());
1347 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1350 case Instruction::Call:
1351 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1352 switch (II->getIntrinsicID()) {
1354 case Intrinsic::bswap: {
1355 // If the only bits demanded come from one byte of the bswap result,
1356 // just shift the input byte into position to eliminate the bswap.
1357 unsigned NLZ = DemandedMask.countLeadingZeros();
1358 unsigned NTZ = DemandedMask.countTrailingZeros();
1360 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
1361 // we need all the bits down to bit 8. Likewise, round NLZ. If we
1362 // have 14 leading zeros, round to 8.
1365 // If we need exactly one byte, we can do this transformation.
1366 if (BitWidth-NLZ-NTZ == 8) {
1367 unsigned ResultBit = NTZ;
1368 unsigned InputBit = BitWidth-NTZ-8;
1370 // Replace this with either a left or right shift to get the byte into
1372 Instruction *NewVal;
1373 if (InputBit > ResultBit)
1374 NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
1375 ConstantInt::get(I->getType(), InputBit-ResultBit));
1377 NewVal = BinaryOperator::CreateShl(I->getOperand(1),
1378 ConstantInt::get(I->getType(), ResultBit-InputBit));
1379 NewVal->takeName(I);
1380 return InsertNewInstBefore(NewVal, *I);
1383 // TODO: Could compute known zero/one bits based on the input.
1388 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1392 // If the client is only demanding bits that we know, return the known
1394 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1395 return ConstantInt::get(RHSKnownOne);
1400 /// SimplifyDemandedVectorElts - The specified value produces a vector with
1401 /// any number of elements. DemandedElts contains the set of elements that are
1402 /// actually used by the caller. This method analyzes which elements of the
1403 /// operand are undef and returns that information in UndefElts.
1405 /// If the information about demanded elements can be used to simplify the
1406 /// operation, the operation is simplified, then the resultant value is
1407 /// returned. This returns null if no change was made.
1408 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
1411 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1412 APInt EltMask(APInt::getAllOnesValue(VWidth));
1413 assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!");
1415 if (isa<UndefValue>(V)) {
1416 // If the entire vector is undefined, just return this info.
1417 UndefElts = EltMask;
1419 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1420 UndefElts = EltMask;
1421 return UndefValue::get(V->getType());
1425 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1426 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1427 Constant *Undef = UndefValue::get(EltTy);
1429 std::vector<Constant*> Elts;
1430 for (unsigned i = 0; i != VWidth; ++i)
1431 if (!DemandedElts[i]) { // If not demanded, set to undef.
1432 Elts.push_back(Undef);
1434 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1435 Elts.push_back(Undef);
1437 } else { // Otherwise, defined.
1438 Elts.push_back(CP->getOperand(i));
1441 // If we changed the constant, return it.
1442 Constant *NewCP = ConstantVector::get(Elts);
1443 return NewCP != CP ? NewCP : 0;
1444 } else if (isa<ConstantAggregateZero>(V)) {
1445 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1448 // Check if this is identity. If so, return 0 since we are not simplifying
1450 if (DemandedElts == ((1ULL << VWidth) -1))
1453 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1454 Constant *Zero = Constant::getNullValue(EltTy);
1455 Constant *Undef = UndefValue::get(EltTy);
1456 std::vector<Constant*> Elts;
1457 for (unsigned i = 0; i != VWidth; ++i) {
1458 Constant *Elt = DemandedElts[i] ? Zero : Undef;
1459 Elts.push_back(Elt);
1461 UndefElts = DemandedElts ^ EltMask;
1462 return ConstantVector::get(Elts);
1465 // Limit search depth.
1469 // If multiple users are using the root value, procede with
1470 // simplification conservatively assuming that all elements
1472 if (!V->hasOneUse()) {
1473 // Quit if we find multiple users of a non-root value though.
1474 // They'll be handled when it's their turn to be visited by
1475 // the main instcombine process.
1477 // TODO: Just compute the UndefElts information recursively.
1480 // Conservatively assume that all elements are needed.
1481 DemandedElts = EltMask;
1484 Instruction *I = dyn_cast<Instruction>(V);
1485 if (!I) return false; // Only analyze instructions.
1487 bool MadeChange = false;
1488 APInt UndefElts2(VWidth, 0);
1490 switch (I->getOpcode()) {
1493 case Instruction::InsertElement: {
1494 // If this is a variable index, we don't know which element it overwrites.
1495 // demand exactly the same input as we produce.
1496 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1498 // Note that we can't propagate undef elt info, because we don't know
1499 // which elt is getting updated.
1500 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1501 UndefElts2, Depth+1);
1502 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1506 // If this is inserting an element that isn't demanded, remove this
1508 unsigned IdxNo = Idx->getZExtValue();
1509 if (IdxNo >= VWidth || !DemandedElts[IdxNo])
1510 return AddSoonDeadInstToWorklist(*I, 0);
1512 // Otherwise, the element inserted overwrites whatever was there, so the
1513 // input demanded set is simpler than the output set.
1514 APInt DemandedElts2 = DemandedElts;
1515 DemandedElts2.clear(IdxNo);
1516 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts2,
1517 UndefElts, Depth+1);
1518 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1520 // The inserted element is defined.
1521 UndefElts.clear(IdxNo);
1524 case Instruction::ShuffleVector: {
1525 ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I);
1526 uint64_t LHSVWidth =
1527 cast<VectorType>(Shuffle->getOperand(0)->getType())->getNumElements();
1528 APInt LeftDemanded(LHSVWidth, 0), RightDemanded(LHSVWidth, 0);
1529 for (unsigned i = 0; i < VWidth; i++) {
1530 if (DemandedElts[i]) {
1531 unsigned MaskVal = Shuffle->getMaskValue(i);
1532 if (MaskVal != -1u) {
1533 assert(MaskVal < LHSVWidth * 2 &&
1534 "shufflevector mask index out of range!");
1535 if (MaskVal < LHSVWidth)
1536 LeftDemanded.set(MaskVal);
1538 RightDemanded.set(MaskVal - LHSVWidth);
1543 APInt UndefElts4(LHSVWidth, 0);
1544 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), LeftDemanded,
1545 UndefElts4, Depth+1);
1546 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1548 APInt UndefElts3(LHSVWidth, 0);
1549 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), RightDemanded,
1550 UndefElts3, Depth+1);
1551 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1553 bool NewUndefElts = false;
1554 for (unsigned i = 0; i < VWidth; i++) {
1555 unsigned MaskVal = Shuffle->getMaskValue(i);
1556 if (MaskVal == -1u) {
1558 } else if (MaskVal < LHSVWidth) {
1559 if (UndefElts4[MaskVal]) {
1560 NewUndefElts = true;
1564 if (UndefElts3[MaskVal - LHSVWidth]) {
1565 NewUndefElts = true;
1572 // Add additional discovered undefs.
1573 std::vector<Constant*> Elts;
1574 for (unsigned i = 0; i < VWidth; ++i) {
1576 Elts.push_back(UndefValue::get(Type::Int32Ty));
1578 Elts.push_back(ConstantInt::get(Type::Int32Ty,
1579 Shuffle->getMaskValue(i)));
1581 I->setOperand(2, ConstantVector::get(Elts));
1586 case Instruction::BitCast: {
1587 // Vector->vector casts only.
1588 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1590 unsigned InVWidth = VTy->getNumElements();
1591 APInt InputDemandedElts(InVWidth, 0);
1594 if (VWidth == InVWidth) {
1595 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1596 // elements as are demanded of us.
1598 InputDemandedElts = DemandedElts;
1599 } else if (VWidth > InVWidth) {
1603 // If there are more elements in the result than there are in the source,
1604 // then an input element is live if any of the corresponding output
1605 // elements are live.
1606 Ratio = VWidth/InVWidth;
1607 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1608 if (DemandedElts[OutIdx])
1609 InputDemandedElts.set(OutIdx/Ratio);
1615 // If there are more elements in the source than there are in the result,
1616 // then an input element is live if the corresponding output element is
1618 Ratio = InVWidth/VWidth;
1619 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1620 if (DemandedElts[InIdx/Ratio])
1621 InputDemandedElts.set(InIdx);
1624 // div/rem demand all inputs, because they don't want divide by zero.
1625 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1626 UndefElts2, Depth+1);
1628 I->setOperand(0, TmpV);
1632 UndefElts = UndefElts2;
1633 if (VWidth > InVWidth) {
1634 assert(0 && "Unimp");
1635 // If there are more elements in the result than there are in the source,
1636 // then an output element is undef if the corresponding input element is
1638 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1639 if (UndefElts2[OutIdx/Ratio])
1640 UndefElts.set(OutIdx);
1641 } else if (VWidth < InVWidth) {
1642 assert(0 && "Unimp");
1643 // If there are more elements in the source than there are in the result,
1644 // then a result element is undef if all of the corresponding input
1645 // elements are undef.
1646 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1647 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1648 if (!UndefElts2[InIdx]) // Not undef?
1649 UndefElts.clear(InIdx/Ratio); // Clear undef bit.
1653 case Instruction::And:
1654 case Instruction::Or:
1655 case Instruction::Xor:
1656 case Instruction::Add:
1657 case Instruction::Sub:
1658 case Instruction::Mul:
1659 // div/rem demand all inputs, because they don't want divide by zero.
1660 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1661 UndefElts, Depth+1);
1662 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1663 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1664 UndefElts2, Depth+1);
1665 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1667 // Output elements are undefined if both are undefined. Consider things
1668 // like undef&0. The result is known zero, not undef.
1669 UndefElts &= UndefElts2;
1672 case Instruction::Call: {
1673 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1675 switch (II->getIntrinsicID()) {
1678 // Binary vector operations that work column-wise. A dest element is a
1679 // function of the corresponding input elements from the two inputs.
1680 case Intrinsic::x86_sse_sub_ss:
1681 case Intrinsic::x86_sse_mul_ss:
1682 case Intrinsic::x86_sse_min_ss:
1683 case Intrinsic::x86_sse_max_ss:
1684 case Intrinsic::x86_sse2_sub_sd:
1685 case Intrinsic::x86_sse2_mul_sd:
1686 case Intrinsic::x86_sse2_min_sd:
1687 case Intrinsic::x86_sse2_max_sd:
1688 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1689 UndefElts, Depth+1);
1690 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1691 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1692 UndefElts2, Depth+1);
1693 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1695 // If only the low elt is demanded and this is a scalarizable intrinsic,
1696 // scalarize it now.
1697 if (DemandedElts == 1) {
1698 switch (II->getIntrinsicID()) {
1700 case Intrinsic::x86_sse_sub_ss:
1701 case Intrinsic::x86_sse_mul_ss:
1702 case Intrinsic::x86_sse2_sub_sd:
1703 case Intrinsic::x86_sse2_mul_sd:
1704 // TODO: Lower MIN/MAX/ABS/etc
1705 Value *LHS = II->getOperand(1);
1706 Value *RHS = II->getOperand(2);
1707 // Extract the element as scalars.
1708 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1709 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1711 switch (II->getIntrinsicID()) {
1712 default: assert(0 && "Case stmts out of sync!");
1713 case Intrinsic::x86_sse_sub_ss:
1714 case Intrinsic::x86_sse2_sub_sd:
1715 TmpV = InsertNewInstBefore(BinaryOperator::CreateSub(LHS, RHS,
1716 II->getName()), *II);
1718 case Intrinsic::x86_sse_mul_ss:
1719 case Intrinsic::x86_sse2_mul_sd:
1720 TmpV = InsertNewInstBefore(BinaryOperator::CreateMul(LHS, RHS,
1721 II->getName()), *II);
1726 InsertElementInst::Create(UndefValue::get(II->getType()), TmpV, 0U,
1728 InsertNewInstBefore(New, *II);
1729 AddSoonDeadInstToWorklist(*II, 0);
1734 // Output elements are undefined if both are undefined. Consider things
1735 // like undef&0. The result is known zero, not undef.
1736 UndefElts &= UndefElts2;
1742 return MadeChange ? I : 0;
1746 /// AssociativeOpt - Perform an optimization on an associative operator. This
1747 /// function is designed to check a chain of associative operators for a
1748 /// potential to apply a certain optimization. Since the optimization may be
1749 /// applicable if the expression was reassociated, this checks the chain, then
1750 /// reassociates the expression as necessary to expose the optimization
1751 /// opportunity. This makes use of a special Functor, which must define
1752 /// 'shouldApply' and 'apply' methods.
1754 template<typename Functor>
1755 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1756 unsigned Opcode = Root.getOpcode();
1757 Value *LHS = Root.getOperand(0);
1759 // Quick check, see if the immediate LHS matches...
1760 if (F.shouldApply(LHS))
1761 return F.apply(Root);
1763 // Otherwise, if the LHS is not of the same opcode as the root, return.
1764 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1765 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1766 // Should we apply this transform to the RHS?
1767 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1769 // If not to the RHS, check to see if we should apply to the LHS...
1770 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1771 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1775 // If the functor wants to apply the optimization to the RHS of LHSI,
1776 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1778 // Now all of the instructions are in the current basic block, go ahead
1779 // and perform the reassociation.
1780 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1782 // First move the selected RHS to the LHS of the root...
1783 Root.setOperand(0, LHSI->getOperand(1));
1785 // Make what used to be the LHS of the root be the user of the root...
1786 Value *ExtraOperand = TmpLHSI->getOperand(1);
1787 if (&Root == TmpLHSI) {
1788 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1791 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1792 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1793 BasicBlock::iterator ARI = &Root; ++ARI;
1794 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
1797 // Now propagate the ExtraOperand down the chain of instructions until we
1799 while (TmpLHSI != LHSI) {
1800 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1801 // Move the instruction to immediately before the chain we are
1802 // constructing to avoid breaking dominance properties.
1803 NextLHSI->moveBefore(ARI);
1806 Value *NextOp = NextLHSI->getOperand(1);
1807 NextLHSI->setOperand(1, ExtraOperand);
1809 ExtraOperand = NextOp;
1812 // Now that the instructions are reassociated, have the functor perform
1813 // the transformation...
1814 return F.apply(Root);
1817 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1824 // AddRHS - Implements: X + X --> X << 1
1827 AddRHS(Value *rhs) : RHS(rhs) {}
1828 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1829 Instruction *apply(BinaryOperator &Add) const {
1830 return BinaryOperator::CreateShl(Add.getOperand(0),
1831 ConstantInt::get(Add.getType(), 1));
1835 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1837 struct AddMaskingAnd {
1839 AddMaskingAnd(Constant *c) : C2(c) {}
1840 bool shouldApply(Value *LHS) const {
1842 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1843 ConstantExpr::getAnd(C1, C2)->isNullValue();
1845 Instruction *apply(BinaryOperator &Add) const {
1846 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
1852 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1854 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1855 return IC->InsertCastBefore(CI->getOpcode(), SO, I.getType(), I);
1858 // Figure out if the constant is the left or the right argument.
1859 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1860 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1862 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1864 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1865 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1868 Value *Op0 = SO, *Op1 = ConstOperand;
1870 std::swap(Op0, Op1);
1872 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1873 New = BinaryOperator::Create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1874 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1875 New = CmpInst::Create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1876 SO->getName()+".cmp");
1878 assert(0 && "Unknown binary instruction type!");
1881 return IC->InsertNewInstBefore(New, I);
1884 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1885 // constant as the other operand, try to fold the binary operator into the
1886 // select arguments. This also works for Cast instructions, which obviously do
1887 // not have a second operand.
1888 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1890 // Don't modify shared select instructions
1891 if (!SI->hasOneUse()) return 0;
1892 Value *TV = SI->getOperand(1);
1893 Value *FV = SI->getOperand(2);
1895 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1896 // Bool selects with constant operands can be folded to logical ops.
1897 if (SI->getType() == Type::Int1Ty) return 0;
1899 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1900 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1902 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
1909 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1910 /// node as operand #0, see if we can fold the instruction into the PHI (which
1911 /// is only possible if all operands to the PHI are constants).
1912 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1913 PHINode *PN = cast<PHINode>(I.getOperand(0));
1914 unsigned NumPHIValues = PN->getNumIncomingValues();
1915 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1917 // Check to see if all of the operands of the PHI are constants. If there is
1918 // one non-constant value, remember the BB it is. If there is more than one
1919 // or if *it* is a PHI, bail out.
1920 BasicBlock *NonConstBB = 0;
1921 for (unsigned i = 0; i != NumPHIValues; ++i)
1922 if (!isa<Constant>(PN->getIncomingValue(i))) {
1923 if (NonConstBB) return 0; // More than one non-const value.
1924 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1925 NonConstBB = PN->getIncomingBlock(i);
1927 // If the incoming non-constant value is in I's block, we have an infinite
1929 if (NonConstBB == I.getParent())
1933 // If there is exactly one non-constant value, we can insert a copy of the
1934 // operation in that block. However, if this is a critical edge, we would be
1935 // inserting the computation one some other paths (e.g. inside a loop). Only
1936 // do this if the pred block is unconditionally branching into the phi block.
1938 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1939 if (!BI || !BI->isUnconditional()) return 0;
1942 // Okay, we can do the transformation: create the new PHI node.
1943 PHINode *NewPN = PHINode::Create(I.getType(), "");
1944 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1945 InsertNewInstBefore(NewPN, *PN);
1946 NewPN->takeName(PN);
1948 // Next, add all of the operands to the PHI.
1949 if (I.getNumOperands() == 2) {
1950 Constant *C = cast<Constant>(I.getOperand(1));
1951 for (unsigned i = 0; i != NumPHIValues; ++i) {
1953 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1954 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1955 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1957 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1959 assert(PN->getIncomingBlock(i) == NonConstBB);
1960 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1961 InV = BinaryOperator::Create(BO->getOpcode(),
1962 PN->getIncomingValue(i), C, "phitmp",
1963 NonConstBB->getTerminator());
1964 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1965 InV = CmpInst::Create(CI->getOpcode(),
1967 PN->getIncomingValue(i), C, "phitmp",
1968 NonConstBB->getTerminator());
1970 assert(0 && "Unknown binop!");
1972 AddToWorkList(cast<Instruction>(InV));
1974 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1977 CastInst *CI = cast<CastInst>(&I);
1978 const Type *RetTy = CI->getType();
1979 for (unsigned i = 0; i != NumPHIValues; ++i) {
1981 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1982 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1984 assert(PN->getIncomingBlock(i) == NonConstBB);
1985 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
1986 I.getType(), "phitmp",
1987 NonConstBB->getTerminator());
1988 AddToWorkList(cast<Instruction>(InV));
1990 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1993 return ReplaceInstUsesWith(I, NewPN);
1997 /// WillNotOverflowSignedAdd - Return true if we can prove that:
1998 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
1999 /// This basically requires proving that the add in the original type would not
2000 /// overflow to change the sign bit or have a carry out.
2001 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
2002 // There are different heuristics we can use for this. Here are some simple
2005 // Add has the property that adding any two 2's complement numbers can only
2006 // have one carry bit which can change a sign. As such, if LHS and RHS each
2007 // have at least two sign bits, we know that the addition of the two values will
2008 // sign extend fine.
2009 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
2013 // If one of the operands only has one non-zero bit, and if the other operand
2014 // has a known-zero bit in a more significant place than it (not including the
2015 // sign bit) the ripple may go up to and fill the zero, but won't change the
2016 // sign. For example, (X & ~4) + 1.
2024 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
2025 bool Changed = SimplifyCommutative(I);
2026 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2028 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2029 // X + undef -> undef
2030 if (isa<UndefValue>(RHS))
2031 return ReplaceInstUsesWith(I, RHS);
2034 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
2035 if (RHSC->isNullValue())
2036 return ReplaceInstUsesWith(I, LHS);
2037 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2038 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2039 (I.getType())->getValueAPF()))
2040 return ReplaceInstUsesWith(I, LHS);
2043 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2044 // X + (signbit) --> X ^ signbit
2045 const APInt& Val = CI->getValue();
2046 uint32_t BitWidth = Val.getBitWidth();
2047 if (Val == APInt::getSignBit(BitWidth))
2048 return BinaryOperator::CreateXor(LHS, RHS);
2050 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2051 // (X & 254)+1 -> (X&254)|1
2052 if (!isa<VectorType>(I.getType()) && SimplifyDemandedInstructionBits(I))
2055 // zext(i1) - 1 -> select i1, 0, -1
2056 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
2057 if (CI->isAllOnesValue() &&
2058 ZI->getOperand(0)->getType() == Type::Int1Ty)
2059 return SelectInst::Create(ZI->getOperand(0),
2060 Constant::getNullValue(I.getType()),
2061 ConstantInt::getAllOnesValue(I.getType()));
2064 if (isa<PHINode>(LHS))
2065 if (Instruction *NV = FoldOpIntoPhi(I))
2068 ConstantInt *XorRHS = 0;
2070 if (isa<ConstantInt>(RHSC) &&
2071 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2072 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
2073 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2075 uint32_t Size = TySizeBits / 2;
2076 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2077 APInt CFF80Val(-C0080Val);
2079 if (TySizeBits > Size) {
2080 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2081 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2082 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2083 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2084 // This is a sign extend if the top bits are known zero.
2085 if (!MaskedValueIsZero(XorLHS,
2086 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2087 Size = 0; // Not a sign ext, but can't be any others either.
2092 C0080Val = APIntOps::lshr(C0080Val, Size);
2093 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2094 } while (Size >= 1);
2096 // FIXME: This shouldn't be necessary. When the backends can handle types
2097 // with funny bit widths then this switch statement should be removed. It
2098 // is just here to get the size of the "middle" type back up to something
2099 // that the back ends can handle.
2100 const Type *MiddleType = 0;
2103 case 32: MiddleType = Type::Int32Ty; break;
2104 case 16: MiddleType = Type::Int16Ty; break;
2105 case 8: MiddleType = Type::Int8Ty; break;
2108 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2109 InsertNewInstBefore(NewTrunc, I);
2110 return new SExtInst(NewTrunc, I.getType(), I.getName());
2115 if (I.getType() == Type::Int1Ty)
2116 return BinaryOperator::CreateXor(LHS, RHS);
2119 if (I.getType()->isInteger()) {
2120 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2122 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2123 if (RHSI->getOpcode() == Instruction::Sub)
2124 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2125 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2127 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2128 if (LHSI->getOpcode() == Instruction::Sub)
2129 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2130 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2135 // -A + -B --> -(A + B)
2136 if (Value *LHSV = dyn_castNegVal(LHS)) {
2137 if (LHS->getType()->isIntOrIntVector()) {
2138 if (Value *RHSV = dyn_castNegVal(RHS)) {
2139 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSV, RHSV, "sum");
2140 InsertNewInstBefore(NewAdd, I);
2141 return BinaryOperator::CreateNeg(NewAdd);
2145 return BinaryOperator::CreateSub(RHS, LHSV);
2149 if (!isa<Constant>(RHS))
2150 if (Value *V = dyn_castNegVal(RHS))
2151 return BinaryOperator::CreateSub(LHS, V);
2155 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2156 if (X == RHS) // X*C + X --> X * (C+1)
2157 return BinaryOperator::CreateMul(RHS, AddOne(C2));
2159 // X*C1 + X*C2 --> X * (C1+C2)
2161 if (X == dyn_castFoldableMul(RHS, C1))
2162 return BinaryOperator::CreateMul(X, Add(C1, C2));
2165 // X + X*C --> X * (C+1)
2166 if (dyn_castFoldableMul(RHS, C2) == LHS)
2167 return BinaryOperator::CreateMul(LHS, AddOne(C2));
2169 // X + ~X --> -1 since ~X = -X-1
2170 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2171 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2174 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2175 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2176 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2179 // A+B --> A|B iff A and B have no bits set in common.
2180 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2181 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2182 APInt LHSKnownOne(IT->getBitWidth(), 0);
2183 APInt LHSKnownZero(IT->getBitWidth(), 0);
2184 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2185 if (LHSKnownZero != 0) {
2186 APInt RHSKnownOne(IT->getBitWidth(), 0);
2187 APInt RHSKnownZero(IT->getBitWidth(), 0);
2188 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2190 // No bits in common -> bitwise or.
2191 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2192 return BinaryOperator::CreateOr(LHS, RHS);
2196 // W*X + Y*Z --> W * (X+Z) iff W == Y
2197 if (I.getType()->isIntOrIntVector()) {
2198 Value *W, *X, *Y, *Z;
2199 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2200 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2204 } else if (Y == X) {
2206 } else if (X == Z) {
2213 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, Z,
2214 LHS->getName()), I);
2215 return BinaryOperator::CreateMul(W, NewAdd);
2220 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2222 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2223 return BinaryOperator::CreateSub(SubOne(CRHS), X);
2225 // (X & FF00) + xx00 -> (X+xx00) & FF00
2226 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2227 Constant *Anded = And(CRHS, C2);
2228 if (Anded == CRHS) {
2229 // See if all bits from the first bit set in the Add RHS up are included
2230 // in the mask. First, get the rightmost bit.
2231 const APInt& AddRHSV = CRHS->getValue();
2233 // Form a mask of all bits from the lowest bit added through the top.
2234 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2236 // See if the and mask includes all of these bits.
2237 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2239 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2240 // Okay, the xform is safe. Insert the new add pronto.
2241 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, CRHS,
2242 LHS->getName()), I);
2243 return BinaryOperator::CreateAnd(NewAdd, C2);
2248 // Try to fold constant add into select arguments.
2249 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2250 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2254 // add (cast *A to intptrtype) B ->
2255 // cast (GEP (cast *A to sbyte*) B) --> intptrtype
2257 CastInst *CI = dyn_cast<CastInst>(LHS);
2260 CI = dyn_cast<CastInst>(RHS);
2263 if (CI && CI->getType()->isSized() &&
2264 (CI->getType()->getPrimitiveSizeInBits() ==
2265 TD->getIntPtrType()->getPrimitiveSizeInBits())
2266 && isa<PointerType>(CI->getOperand(0)->getType())) {
2268 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2269 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2270 PointerType::get(Type::Int8Ty, AS), I);
2271 I2 = InsertNewInstBefore(GetElementPtrInst::Create(I2, Other, "ctg2"), I);
2272 return new PtrToIntInst(I2, CI->getType());
2276 // add (select X 0 (sub n A)) A --> select X A n
2278 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2281 SI = dyn_cast<SelectInst>(RHS);
2284 if (SI && SI->hasOneUse()) {
2285 Value *TV = SI->getTrueValue();
2286 Value *FV = SI->getFalseValue();
2289 // Can we fold the add into the argument of the select?
2290 // We check both true and false select arguments for a matching subtract.
2291 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Specific(A))))
2292 // Fold the add into the true select value.
2293 return SelectInst::Create(SI->getCondition(), N, A);
2294 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Specific(A))))
2295 // Fold the add into the false select value.
2296 return SelectInst::Create(SI->getCondition(), A, N);
2300 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2301 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2302 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2303 return ReplaceInstUsesWith(I, LHS);
2305 // Check for (add (sext x), y), see if we can merge this into an
2306 // integer add followed by a sext.
2307 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2308 // (add (sext x), cst) --> (sext (add x, cst'))
2309 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2311 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2312 if (LHSConv->hasOneUse() &&
2313 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2314 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2315 // Insert the new, smaller add.
2316 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2318 InsertNewInstBefore(NewAdd, I);
2319 return new SExtInst(NewAdd, I.getType());
2323 // (add (sext x), (sext y)) --> (sext (add int x, y))
2324 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2325 // Only do this if x/y have the same type, if at last one of them has a
2326 // single use (so we don't increase the number of sexts), and if the
2327 // integer add will not overflow.
2328 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2329 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2330 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2331 RHSConv->getOperand(0))) {
2332 // Insert the new integer add.
2333 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2334 RHSConv->getOperand(0),
2336 InsertNewInstBefore(NewAdd, I);
2337 return new SExtInst(NewAdd, I.getType());
2342 // Check for (add double (sitofp x), y), see if we can merge this into an
2343 // integer add followed by a promotion.
2344 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2345 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2346 // ... if the constant fits in the integer value. This is useful for things
2347 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2348 // requires a constant pool load, and generally allows the add to be better
2350 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2352 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2353 if (LHSConv->hasOneUse() &&
2354 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2355 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2356 // Insert the new integer add.
2357 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2359 InsertNewInstBefore(NewAdd, I);
2360 return new SIToFPInst(NewAdd, I.getType());
2364 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2365 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2366 // Only do this if x/y have the same type, if at last one of them has a
2367 // single use (so we don't increase the number of int->fp conversions),
2368 // and if the integer add will not overflow.
2369 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2370 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2371 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2372 RHSConv->getOperand(0))) {
2373 // Insert the new integer add.
2374 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2375 RHSConv->getOperand(0),
2377 InsertNewInstBefore(NewAdd, I);
2378 return new SIToFPInst(NewAdd, I.getType());
2383 return Changed ? &I : 0;
2386 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2387 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2389 if (Op0 == Op1 && // sub X, X -> 0
2390 !I.getType()->isFPOrFPVector())
2391 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2393 // If this is a 'B = x-(-A)', change to B = x+A...
2394 if (Value *V = dyn_castNegVal(Op1))
2395 return BinaryOperator::CreateAdd(Op0, V);
2397 if (isa<UndefValue>(Op0))
2398 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2399 if (isa<UndefValue>(Op1))
2400 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2402 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2403 // Replace (-1 - A) with (~A)...
2404 if (C->isAllOnesValue())
2405 return BinaryOperator::CreateNot(Op1);
2407 // C - ~X == X + (1+C)
2409 if (match(Op1, m_Not(m_Value(X))))
2410 return BinaryOperator::CreateAdd(X, AddOne(C));
2412 // -(X >>u 31) -> (X >>s 31)
2413 // -(X >>s 31) -> (X >>u 31)
2415 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2416 if (SI->getOpcode() == Instruction::LShr) {
2417 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2418 // Check to see if we are shifting out everything but the sign bit.
2419 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2420 SI->getType()->getPrimitiveSizeInBits()-1) {
2421 // Ok, the transformation is safe. Insert AShr.
2422 return BinaryOperator::Create(Instruction::AShr,
2423 SI->getOperand(0), CU, SI->getName());
2427 else if (SI->getOpcode() == Instruction::AShr) {
2428 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2429 // Check to see if we are shifting out everything but the sign bit.
2430 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2431 SI->getType()->getPrimitiveSizeInBits()-1) {
2432 // Ok, the transformation is safe. Insert LShr.
2433 return BinaryOperator::CreateLShr(
2434 SI->getOperand(0), CU, SI->getName());
2441 // Try to fold constant sub into select arguments.
2442 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2443 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2447 if (I.getType() == Type::Int1Ty)
2448 return BinaryOperator::CreateXor(Op0, Op1);
2450 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2451 if (Op1I->getOpcode() == Instruction::Add &&
2452 !Op0->getType()->isFPOrFPVector()) {
2453 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2454 return BinaryOperator::CreateNeg(Op1I->getOperand(1), I.getName());
2455 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2456 return BinaryOperator::CreateNeg(Op1I->getOperand(0), I.getName());
2457 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2458 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2459 // C1-(X+C2) --> (C1-C2)-X
2460 return BinaryOperator::CreateSub(Subtract(CI1, CI2),
2461 Op1I->getOperand(0));
2465 if (Op1I->hasOneUse()) {
2466 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2467 // is not used by anyone else...
2469 if (Op1I->getOpcode() == Instruction::Sub &&
2470 !Op1I->getType()->isFPOrFPVector()) {
2471 // Swap the two operands of the subexpr...
2472 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2473 Op1I->setOperand(0, IIOp1);
2474 Op1I->setOperand(1, IIOp0);
2476 // Create the new top level add instruction...
2477 return BinaryOperator::CreateAdd(Op0, Op1);
2480 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2482 if (Op1I->getOpcode() == Instruction::And &&
2483 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2484 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2487 InsertNewInstBefore(BinaryOperator::CreateNot(OtherOp, "B.not"), I);
2488 return BinaryOperator::CreateAnd(Op0, NewNot);
2491 // 0 - (X sdiv C) -> (X sdiv -C)
2492 if (Op1I->getOpcode() == Instruction::SDiv)
2493 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2495 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2496 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2497 ConstantExpr::getNeg(DivRHS));
2499 // X - X*C --> X * (1-C)
2500 ConstantInt *C2 = 0;
2501 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2502 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2503 return BinaryOperator::CreateMul(Op0, CP1);
2508 if (!Op0->getType()->isFPOrFPVector())
2509 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2510 if (Op0I->getOpcode() == Instruction::Add) {
2511 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2512 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2513 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2514 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2515 } else if (Op0I->getOpcode() == Instruction::Sub) {
2516 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2517 return BinaryOperator::CreateNeg(Op0I->getOperand(1), I.getName());
2522 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2523 if (X == Op1) // X*C - X --> X * (C-1)
2524 return BinaryOperator::CreateMul(Op1, SubOne(C1));
2526 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2527 if (X == dyn_castFoldableMul(Op1, C2))
2528 return BinaryOperator::CreateMul(X, Subtract(C1, C2));
2533 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2534 /// comparison only checks the sign bit. If it only checks the sign bit, set
2535 /// TrueIfSigned if the result of the comparison is true when the input value is
2537 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2538 bool &TrueIfSigned) {
2540 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2541 TrueIfSigned = true;
2542 return RHS->isZero();
2543 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2544 TrueIfSigned = true;
2545 return RHS->isAllOnesValue();
2546 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2547 TrueIfSigned = false;
2548 return RHS->isAllOnesValue();
2549 case ICmpInst::ICMP_UGT:
2550 // True if LHS u> RHS and RHS == high-bit-mask - 1
2551 TrueIfSigned = true;
2552 return RHS->getValue() ==
2553 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2554 case ICmpInst::ICMP_UGE:
2555 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2556 TrueIfSigned = true;
2557 return RHS->getValue().isSignBit();
2563 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2564 bool Changed = SimplifyCommutative(I);
2565 Value *Op0 = I.getOperand(0);
2567 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2568 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2570 // Simplify mul instructions with a constant RHS...
2571 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2572 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2574 // ((X << C1)*C2) == (X * (C2 << C1))
2575 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2576 if (SI->getOpcode() == Instruction::Shl)
2577 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2578 return BinaryOperator::CreateMul(SI->getOperand(0),
2579 ConstantExpr::getShl(CI, ShOp));
2582 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2583 if (CI->equalsInt(1)) // X * 1 == X
2584 return ReplaceInstUsesWith(I, Op0);
2585 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2586 return BinaryOperator::CreateNeg(Op0, I.getName());
2588 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2589 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2590 return BinaryOperator::CreateShl(Op0,
2591 ConstantInt::get(Op0->getType(), Val.logBase2()));
2593 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2594 if (Op1F->isNullValue())
2595 return ReplaceInstUsesWith(I, Op1);
2597 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2598 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2599 if (Op1F->isExactlyValue(1.0))
2600 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2601 } else if (isa<VectorType>(Op1->getType())) {
2602 if (isa<ConstantAggregateZero>(Op1))
2603 return ReplaceInstUsesWith(I, Op1);
2605 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2606 if (Op1V->isAllOnesValue()) // X * -1 == 0 - X
2607 return BinaryOperator::CreateNeg(Op0, I.getName());
2609 // As above, vector X*splat(1.0) -> X in all defined cases.
2610 if (Constant *Splat = Op1V->getSplatValue()) {
2611 if (ConstantFP *F = dyn_cast<ConstantFP>(Splat))
2612 if (F->isExactlyValue(1.0))
2613 return ReplaceInstUsesWith(I, Op0);
2614 if (ConstantInt *CI = dyn_cast<ConstantInt>(Splat))
2615 if (CI->equalsInt(1))
2616 return ReplaceInstUsesWith(I, Op0);
2621 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2622 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2623 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1)) {
2624 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2625 Instruction *Add = BinaryOperator::CreateMul(Op0I->getOperand(0),
2627 InsertNewInstBefore(Add, I);
2628 Value *C1C2 = ConstantExpr::getMul(Op1,
2629 cast<Constant>(Op0I->getOperand(1)));
2630 return BinaryOperator::CreateAdd(Add, C1C2);
2634 // Try to fold constant mul into select arguments.
2635 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2636 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2639 if (isa<PHINode>(Op0))
2640 if (Instruction *NV = FoldOpIntoPhi(I))
2644 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2645 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2646 return BinaryOperator::CreateMul(Op0v, Op1v);
2648 // (X / Y) * Y = X - (X % Y)
2649 // (X / Y) * -Y = (X % Y) - X
2651 Value *Op1 = I.getOperand(1);
2652 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0);
2654 (BO->getOpcode() != Instruction::UDiv &&
2655 BO->getOpcode() != Instruction::SDiv)) {
2657 BO = dyn_cast<BinaryOperator>(I.getOperand(1));
2659 Value *Neg = dyn_castNegVal(Op1);
2660 if (BO && BO->hasOneUse() &&
2661 (BO->getOperand(1) == Op1 || BO->getOperand(1) == Neg) &&
2662 (BO->getOpcode() == Instruction::UDiv ||
2663 BO->getOpcode() == Instruction::SDiv)) {
2664 Value *Op0BO = BO->getOperand(0), *Op1BO = BO->getOperand(1);
2667 if (BO->getOpcode() == Instruction::UDiv)
2668 Rem = BinaryOperator::CreateURem(Op0BO, Op1BO);
2670 Rem = BinaryOperator::CreateSRem(Op0BO, Op1BO);
2672 InsertNewInstBefore(Rem, I);
2676 return BinaryOperator::CreateSub(Op0BO, Rem);
2678 return BinaryOperator::CreateSub(Rem, Op0BO);
2682 if (I.getType() == Type::Int1Ty)
2683 return BinaryOperator::CreateAnd(Op0, I.getOperand(1));
2685 // If one of the operands of the multiply is a cast from a boolean value, then
2686 // we know the bool is either zero or one, so this is a 'masking' multiply.
2687 // See if we can simplify things based on how the boolean was originally
2689 CastInst *BoolCast = 0;
2690 if (ZExtInst *CI = dyn_cast<ZExtInst>(Op0))
2691 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2694 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2695 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2698 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2699 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2700 const Type *SCOpTy = SCIOp0->getType();
2703 // If the icmp is true iff the sign bit of X is set, then convert this
2704 // multiply into a shift/and combination.
2705 if (isa<ConstantInt>(SCIOp1) &&
2706 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2708 // Shift the X value right to turn it into "all signbits".
2709 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2710 SCOpTy->getPrimitiveSizeInBits()-1);
2712 InsertNewInstBefore(
2713 BinaryOperator::Create(Instruction::AShr, SCIOp0, Amt,
2714 BoolCast->getOperand(0)->getName()+
2717 // If the multiply type is not the same as the source type, sign extend
2718 // or truncate to the multiply type.
2719 if (I.getType() != V->getType()) {
2720 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2721 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2722 Instruction::CastOps opcode =
2723 (SrcBits == DstBits ? Instruction::BitCast :
2724 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2725 V = InsertCastBefore(opcode, V, I.getType(), I);
2728 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2729 return BinaryOperator::CreateAnd(V, OtherOp);
2734 return Changed ? &I : 0;
2737 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
2739 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
2740 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
2742 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
2743 int NonNullOperand = -1;
2744 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2745 if (ST->isNullValue())
2747 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
2748 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2749 if (ST->isNullValue())
2752 if (NonNullOperand == -1)
2755 Value *SelectCond = SI->getOperand(0);
2757 // Change the div/rem to use 'Y' instead of the select.
2758 I.setOperand(1, SI->getOperand(NonNullOperand));
2760 // Okay, we know we replace the operand of the div/rem with 'Y' with no
2761 // problem. However, the select, or the condition of the select may have
2762 // multiple uses. Based on our knowledge that the operand must be non-zero,
2763 // propagate the known value for the select into other uses of it, and
2764 // propagate a known value of the condition into its other users.
2766 // If the select and condition only have a single use, don't bother with this,
2768 if (SI->use_empty() && SelectCond->hasOneUse())
2771 // Scan the current block backward, looking for other uses of SI.
2772 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
2774 while (BBI != BBFront) {
2776 // If we found a call to a function, we can't assume it will return, so
2777 // information from below it cannot be propagated above it.
2778 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
2781 // Replace uses of the select or its condition with the known values.
2782 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
2785 *I = SI->getOperand(NonNullOperand);
2787 } else if (*I == SelectCond) {
2788 *I = NonNullOperand == 1 ? ConstantInt::getTrue() :
2789 ConstantInt::getFalse();
2794 // If we past the instruction, quit looking for it.
2797 if (&*BBI == SelectCond)
2800 // If we ran out of things to eliminate, break out of the loop.
2801 if (SelectCond == 0 && SI == 0)
2809 /// This function implements the transforms on div instructions that work
2810 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2811 /// used by the visitors to those instructions.
2812 /// @brief Transforms common to all three div instructions
2813 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2814 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2816 // undef / X -> 0 for integer.
2817 // undef / X -> undef for FP (the undef could be a snan).
2818 if (isa<UndefValue>(Op0)) {
2819 if (Op0->getType()->isFPOrFPVector())
2820 return ReplaceInstUsesWith(I, Op0);
2821 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2824 // X / undef -> undef
2825 if (isa<UndefValue>(Op1))
2826 return ReplaceInstUsesWith(I, Op1);
2831 /// This function implements the transforms common to both integer division
2832 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2833 /// division instructions.
2834 /// @brief Common integer divide transforms
2835 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2836 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2838 // (sdiv X, X) --> 1 (udiv X, X) --> 1
2840 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
2841 ConstantInt *CI = ConstantInt::get(Ty->getElementType(), 1);
2842 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
2843 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
2846 ConstantInt *CI = ConstantInt::get(I.getType(), 1);
2847 return ReplaceInstUsesWith(I, CI);
2850 if (Instruction *Common = commonDivTransforms(I))
2853 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2854 // This does not apply for fdiv.
2855 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
2858 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2860 if (RHS->equalsInt(1))
2861 return ReplaceInstUsesWith(I, Op0);
2863 // (X / C1) / C2 -> X / (C1*C2)
2864 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2865 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2866 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2867 if (MultiplyOverflows(RHS, LHSRHS, I.getOpcode()==Instruction::SDiv))
2868 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2870 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
2871 Multiply(RHS, LHSRHS));
2874 if (!RHS->isZero()) { // avoid X udiv 0
2875 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2876 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2878 if (isa<PHINode>(Op0))
2879 if (Instruction *NV = FoldOpIntoPhi(I))
2884 // 0 / X == 0, we don't need to preserve faults!
2885 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2886 if (LHS->equalsInt(0))
2887 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2889 // It can't be division by zero, hence it must be division by one.
2890 if (I.getType() == Type::Int1Ty)
2891 return ReplaceInstUsesWith(I, Op0);
2893 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2894 if (ConstantInt *X = cast_or_null<ConstantInt>(Op1V->getSplatValue()))
2897 return ReplaceInstUsesWith(I, Op0);
2903 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2904 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2906 // Handle the integer div common cases
2907 if (Instruction *Common = commonIDivTransforms(I))
2910 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2911 // X udiv C^2 -> X >> C
2912 // Check to see if this is an unsigned division with an exact power of 2,
2913 // if so, convert to a right shift.
2914 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2915 return BinaryOperator::CreateLShr(Op0,
2916 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2918 // X udiv C, where C >= signbit
2919 if (C->getValue().isNegative()) {
2920 Value *IC = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_ULT, Op0, C),
2922 return SelectInst::Create(IC, Constant::getNullValue(I.getType()),
2923 ConstantInt::get(I.getType(), 1));
2927 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2928 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2929 if (RHSI->getOpcode() == Instruction::Shl &&
2930 isa<ConstantInt>(RHSI->getOperand(0))) {
2931 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2932 if (C1.isPowerOf2()) {
2933 Value *N = RHSI->getOperand(1);
2934 const Type *NTy = N->getType();
2935 if (uint32_t C2 = C1.logBase2()) {
2936 Constant *C2V = ConstantInt::get(NTy, C2);
2937 N = InsertNewInstBefore(BinaryOperator::CreateAdd(N, C2V, "tmp"), I);
2939 return BinaryOperator::CreateLShr(Op0, N);
2944 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2945 // where C1&C2 are powers of two.
2946 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2947 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2948 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2949 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2950 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2951 // Compute the shift amounts
2952 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2953 // Construct the "on true" case of the select
2954 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2955 Instruction *TSI = BinaryOperator::CreateLShr(
2956 Op0, TC, SI->getName()+".t");
2957 TSI = InsertNewInstBefore(TSI, I);
2959 // Construct the "on false" case of the select
2960 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2961 Instruction *FSI = BinaryOperator::CreateLShr(
2962 Op0, FC, SI->getName()+".f");
2963 FSI = InsertNewInstBefore(FSI, I);
2965 // construct the select instruction and return it.
2966 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
2972 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2973 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2975 // Handle the integer div common cases
2976 if (Instruction *Common = commonIDivTransforms(I))
2979 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2981 if (RHS->isAllOnesValue())
2982 return BinaryOperator::CreateNeg(Op0);
2985 // If the sign bits of both operands are zero (i.e. we can prove they are
2986 // unsigned inputs), turn this into a udiv.
2987 if (I.getType()->isInteger()) {
2988 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2989 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2990 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2991 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
2998 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2999 return commonDivTransforms(I);
3002 /// This function implements the transforms on rem instructions that work
3003 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
3004 /// is used by the visitors to those instructions.
3005 /// @brief Transforms common to all three rem instructions
3006 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
3007 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3009 if (isa<UndefValue>(Op0)) { // undef % X -> 0
3010 if (I.getType()->isFPOrFPVector())
3011 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
3012 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3014 if (isa<UndefValue>(Op1))
3015 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
3017 // Handle cases involving: rem X, (select Cond, Y, Z)
3018 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
3024 /// This function implements the transforms common to both integer remainder
3025 /// instructions (urem and srem). It is called by the visitors to those integer
3026 /// remainder instructions.
3027 /// @brief Common integer remainder transforms
3028 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3029 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3031 if (Instruction *common = commonRemTransforms(I))
3034 // 0 % X == 0 for integer, we don't need to preserve faults!
3035 if (Constant *LHS = dyn_cast<Constant>(Op0))
3036 if (LHS->isNullValue())
3037 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3039 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3040 // X % 0 == undef, we don't need to preserve faults!
3041 if (RHS->equalsInt(0))
3042 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3044 if (RHS->equalsInt(1)) // X % 1 == 0
3045 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3047 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3048 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3049 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3051 } else if (isa<PHINode>(Op0I)) {
3052 if (Instruction *NV = FoldOpIntoPhi(I))
3056 // See if we can fold away this rem instruction.
3057 if (SimplifyDemandedInstructionBits(I))
3065 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3066 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3068 if (Instruction *common = commonIRemTransforms(I))
3071 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3072 // X urem C^2 -> X and C
3073 // Check to see if this is an unsigned remainder with an exact power of 2,
3074 // if so, convert to a bitwise and.
3075 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3076 if (C->getValue().isPowerOf2())
3077 return BinaryOperator::CreateAnd(Op0, SubOne(C));
3080 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3081 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3082 if (RHSI->getOpcode() == Instruction::Shl &&
3083 isa<ConstantInt>(RHSI->getOperand(0))) {
3084 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3085 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
3086 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
3088 return BinaryOperator::CreateAnd(Op0, Add);
3093 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3094 // where C1&C2 are powers of two.
3095 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3096 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3097 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3098 // STO == 0 and SFO == 0 handled above.
3099 if ((STO->getValue().isPowerOf2()) &&
3100 (SFO->getValue().isPowerOf2())) {
3101 Value *TrueAnd = InsertNewInstBefore(
3102 BinaryOperator::CreateAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
3103 Value *FalseAnd = InsertNewInstBefore(
3104 BinaryOperator::CreateAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
3105 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3113 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3114 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3116 // Handle the integer rem common cases
3117 if (Instruction *common = commonIRemTransforms(I))
3120 if (Value *RHSNeg = dyn_castNegVal(Op1))
3121 if (!isa<Constant>(RHSNeg) ||
3122 (isa<ConstantInt>(RHSNeg) &&
3123 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
3125 AddUsesToWorkList(I);
3126 I.setOperand(1, RHSNeg);
3130 // If the sign bits of both operands are zero (i.e. we can prove they are
3131 // unsigned inputs), turn this into a urem.
3132 if (I.getType()->isInteger()) {
3133 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3134 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3135 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3136 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3140 // If it's a constant vector, flip any negative values positive.
3141 if (ConstantVector *RHSV = dyn_cast<ConstantVector>(Op1)) {
3142 unsigned VWidth = RHSV->getNumOperands();
3144 bool hasNegative = false;
3145 for (unsigned i = 0; !hasNegative && i != VWidth; ++i)
3146 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i)))
3147 if (RHS->getValue().isNegative())
3151 std::vector<Constant *> Elts(VWidth);
3152 for (unsigned i = 0; i != VWidth; ++i) {
3153 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i))) {
3154 if (RHS->getValue().isNegative())
3155 Elts[i] = cast<ConstantInt>(ConstantExpr::getNeg(RHS));
3161 Constant *NewRHSV = ConstantVector::get(Elts);
3162 if (NewRHSV != RHSV) {
3163 AddUsesToWorkList(I);
3164 I.setOperand(1, NewRHSV);
3173 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3174 return commonRemTransforms(I);
3177 // isOneBitSet - Return true if there is exactly one bit set in the specified
3179 static bool isOneBitSet(const ConstantInt *CI) {
3180 return CI->getValue().isPowerOf2();
3183 // isHighOnes - Return true if the constant is of the form 1+0+.
3184 // This is the same as lowones(~X).
3185 static bool isHighOnes(const ConstantInt *CI) {
3186 return (~CI->getValue() + 1).isPowerOf2();
3189 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3190 /// are carefully arranged to allow folding of expressions such as:
3192 /// (A < B) | (A > B) --> (A != B)
3194 /// Note that this is only valid if the first and second predicates have the
3195 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3197 /// Three bits are used to represent the condition, as follows:
3202 /// <=> Value Definition
3203 /// 000 0 Always false
3210 /// 111 7 Always true
3212 static unsigned getICmpCode(const ICmpInst *ICI) {
3213 switch (ICI->getPredicate()) {
3215 case ICmpInst::ICMP_UGT: return 1; // 001
3216 case ICmpInst::ICMP_SGT: return 1; // 001
3217 case ICmpInst::ICMP_EQ: return 2; // 010
3218 case ICmpInst::ICMP_UGE: return 3; // 011
3219 case ICmpInst::ICMP_SGE: return 3; // 011
3220 case ICmpInst::ICMP_ULT: return 4; // 100
3221 case ICmpInst::ICMP_SLT: return 4; // 100
3222 case ICmpInst::ICMP_NE: return 5; // 101
3223 case ICmpInst::ICMP_ULE: return 6; // 110
3224 case ICmpInst::ICMP_SLE: return 6; // 110
3227 assert(0 && "Invalid ICmp predicate!");
3232 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
3233 /// predicate into a three bit mask. It also returns whether it is an ordered
3234 /// predicate by reference.
3235 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
3238 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
3239 case FCmpInst::FCMP_UNO: return 0; // 000
3240 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
3241 case FCmpInst::FCMP_UGT: return 1; // 001
3242 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
3243 case FCmpInst::FCMP_UEQ: return 2; // 010
3244 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
3245 case FCmpInst::FCMP_UGE: return 3; // 011
3246 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
3247 case FCmpInst::FCMP_ULT: return 4; // 100
3248 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
3249 case FCmpInst::FCMP_UNE: return 5; // 101
3250 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
3251 case FCmpInst::FCMP_ULE: return 6; // 110
3254 // Not expecting FCMP_FALSE and FCMP_TRUE;
3255 assert(0 && "Unexpected FCmp predicate!");
3260 /// getICmpValue - This is the complement of getICmpCode, which turns an
3261 /// opcode and two operands into either a constant true or false, or a brand
3262 /// new ICmp instruction. The sign is passed in to determine which kind
3263 /// of predicate to use in the new icmp instruction.
3264 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3266 default: assert(0 && "Illegal ICmp code!");
3267 case 0: return ConstantInt::getFalse();
3270 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3272 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3273 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3276 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3278 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3281 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3283 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3284 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3287 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3289 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3290 case 7: return ConstantInt::getTrue();
3294 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
3295 /// opcode and two operands into either a FCmp instruction. isordered is passed
3296 /// in to determine which kind of predicate to use in the new fcmp instruction.
3297 static Value *getFCmpValue(bool isordered, unsigned code,
3298 Value *LHS, Value *RHS) {
3300 default: assert(0 && "Illegal FCmp code!");
3303 return new FCmpInst(FCmpInst::FCMP_ORD, LHS, RHS);
3305 return new FCmpInst(FCmpInst::FCMP_UNO, LHS, RHS);
3308 return new FCmpInst(FCmpInst::FCMP_OGT, LHS, RHS);
3310 return new FCmpInst(FCmpInst::FCMP_UGT, LHS, RHS);
3313 return new FCmpInst(FCmpInst::FCMP_OEQ, LHS, RHS);
3315 return new FCmpInst(FCmpInst::FCMP_UEQ, LHS, RHS);
3318 return new FCmpInst(FCmpInst::FCMP_OGE, LHS, RHS);
3320 return new FCmpInst(FCmpInst::FCMP_UGE, LHS, RHS);
3323 return new FCmpInst(FCmpInst::FCMP_OLT, LHS, RHS);
3325 return new FCmpInst(FCmpInst::FCMP_ULT, LHS, RHS);
3328 return new FCmpInst(FCmpInst::FCMP_ONE, LHS, RHS);
3330 return new FCmpInst(FCmpInst::FCMP_UNE, LHS, RHS);
3333 return new FCmpInst(FCmpInst::FCMP_OLE, LHS, RHS);
3335 return new FCmpInst(FCmpInst::FCMP_ULE, LHS, RHS);
3336 case 7: return ConstantInt::getTrue();
3340 /// PredicatesFoldable - Return true if both predicates match sign or if at
3341 /// least one of them is an equality comparison (which is signless).
3342 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3343 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3344 (ICmpInst::isSignedPredicate(p1) && ICmpInst::isEquality(p2)) ||
3345 (ICmpInst::isSignedPredicate(p2) && ICmpInst::isEquality(p1));
3349 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3350 struct FoldICmpLogical {
3353 ICmpInst::Predicate pred;
3354 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3355 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3356 pred(ICI->getPredicate()) {}
3357 bool shouldApply(Value *V) const {
3358 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3359 if (PredicatesFoldable(pred, ICI->getPredicate()))
3360 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3361 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3364 Instruction *apply(Instruction &Log) const {
3365 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3366 if (ICI->getOperand(0) != LHS) {
3367 assert(ICI->getOperand(1) == LHS);
3368 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3371 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3372 unsigned LHSCode = getICmpCode(ICI);
3373 unsigned RHSCode = getICmpCode(RHSICI);
3375 switch (Log.getOpcode()) {
3376 case Instruction::And: Code = LHSCode & RHSCode; break;
3377 case Instruction::Or: Code = LHSCode | RHSCode; break;
3378 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3379 default: assert(0 && "Illegal logical opcode!"); return 0;
3382 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3383 ICmpInst::isSignedPredicate(ICI->getPredicate());
3385 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3386 if (Instruction *I = dyn_cast<Instruction>(RV))
3388 // Otherwise, it's a constant boolean value...
3389 return IC.ReplaceInstUsesWith(Log, RV);
3392 } // end anonymous namespace
3394 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3395 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3396 // guaranteed to be a binary operator.
3397 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3399 ConstantInt *AndRHS,
3400 BinaryOperator &TheAnd) {
3401 Value *X = Op->getOperand(0);
3402 Constant *Together = 0;
3404 Together = And(AndRHS, OpRHS);
3406 switch (Op->getOpcode()) {
3407 case Instruction::Xor:
3408 if (Op->hasOneUse()) {
3409 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3410 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3411 InsertNewInstBefore(And, TheAnd);
3413 return BinaryOperator::CreateXor(And, Together);
3416 case Instruction::Or:
3417 if (Together == AndRHS) // (X | C) & C --> C
3418 return ReplaceInstUsesWith(TheAnd, AndRHS);
3420 if (Op->hasOneUse() && Together != OpRHS) {
3421 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3422 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3423 InsertNewInstBefore(Or, TheAnd);
3425 return BinaryOperator::CreateAnd(Or, AndRHS);
3428 case Instruction::Add:
3429 if (Op->hasOneUse()) {
3430 // Adding a one to a single bit bit-field should be turned into an XOR
3431 // of the bit. First thing to check is to see if this AND is with a
3432 // single bit constant.
3433 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3435 // If there is only one bit set...
3436 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3437 // Ok, at this point, we know that we are masking the result of the
3438 // ADD down to exactly one bit. If the constant we are adding has
3439 // no bits set below this bit, then we can eliminate the ADD.
3440 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3442 // Check to see if any bits below the one bit set in AndRHSV are set.
3443 if ((AddRHS & (AndRHSV-1)) == 0) {
3444 // If not, the only thing that can effect the output of the AND is
3445 // the bit specified by AndRHSV. If that bit is set, the effect of
3446 // the XOR is to toggle the bit. If it is clear, then the ADD has
3448 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3449 TheAnd.setOperand(0, X);
3452 // Pull the XOR out of the AND.
3453 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3454 InsertNewInstBefore(NewAnd, TheAnd);
3455 NewAnd->takeName(Op);
3456 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3463 case Instruction::Shl: {
3464 // We know that the AND will not produce any of the bits shifted in, so if
3465 // the anded constant includes them, clear them now!
3467 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3468 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3469 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3470 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3472 if (CI->getValue() == ShlMask) {
3473 // Masking out bits that the shift already masks
3474 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3475 } else if (CI != AndRHS) { // Reducing bits set in and.
3476 TheAnd.setOperand(1, CI);
3481 case Instruction::LShr:
3483 // We know that the AND will not produce any of the bits shifted in, so if
3484 // the anded constant includes them, clear them now! This only applies to
3485 // unsigned shifts, because a signed shr may bring in set bits!
3487 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3488 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3489 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3490 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3492 if (CI->getValue() == ShrMask) {
3493 // Masking out bits that the shift already masks.
3494 return ReplaceInstUsesWith(TheAnd, Op);
3495 } else if (CI != AndRHS) {
3496 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3501 case Instruction::AShr:
3503 // See if this is shifting in some sign extension, then masking it out
3505 if (Op->hasOneUse()) {
3506 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3507 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3508 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3509 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3510 if (C == AndRHS) { // Masking out bits shifted in.
3511 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3512 // Make the argument unsigned.
3513 Value *ShVal = Op->getOperand(0);
3514 ShVal = InsertNewInstBefore(
3515 BinaryOperator::CreateLShr(ShVal, OpRHS,
3516 Op->getName()), TheAnd);
3517 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3526 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3527 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3528 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3529 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3530 /// insert new instructions.
3531 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3532 bool isSigned, bool Inside,
3534 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3535 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3536 "Lo is not <= Hi in range emission code!");
3539 if (Lo == Hi) // Trivially false.
3540 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3542 // V >= Min && V < Hi --> V < Hi
3543 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3544 ICmpInst::Predicate pred = (isSigned ?
3545 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3546 return new ICmpInst(pred, V, Hi);
3549 // Emit V-Lo <u Hi-Lo
3550 Constant *NegLo = ConstantExpr::getNeg(Lo);
3551 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3552 InsertNewInstBefore(Add, IB);
3553 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3554 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3557 if (Lo == Hi) // Trivially true.
3558 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3560 // V < Min || V >= Hi -> V > Hi-1
3561 Hi = SubOne(cast<ConstantInt>(Hi));
3562 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3563 ICmpInst::Predicate pred = (isSigned ?
3564 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3565 return new ICmpInst(pred, V, Hi);
3568 // Emit V-Lo >u Hi-1-Lo
3569 // Note that Hi has already had one subtracted from it, above.
3570 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3571 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3572 InsertNewInstBefore(Add, IB);
3573 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3574 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3577 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3578 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3579 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3580 // not, since all 1s are not contiguous.
3581 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3582 const APInt& V = Val->getValue();
3583 uint32_t BitWidth = Val->getType()->getBitWidth();
3584 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3586 // look for the first zero bit after the run of ones
3587 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3588 // look for the first non-zero bit
3589 ME = V.getActiveBits();
3593 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3594 /// where isSub determines whether the operator is a sub. If we can fold one of
3595 /// the following xforms:
3597 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3598 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3599 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3601 /// return (A +/- B).
3603 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3604 ConstantInt *Mask, bool isSub,
3606 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3607 if (!LHSI || LHSI->getNumOperands() != 2 ||
3608 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3610 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3612 switch (LHSI->getOpcode()) {
3614 case Instruction::And:
3615 if (And(N, Mask) == Mask) {
3616 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3617 if ((Mask->getValue().countLeadingZeros() +
3618 Mask->getValue().countPopulation()) ==
3619 Mask->getValue().getBitWidth())
3622 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3623 // part, we don't need any explicit masks to take them out of A. If that
3624 // is all N is, ignore it.
3625 uint32_t MB = 0, ME = 0;
3626 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3627 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3628 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3629 if (MaskedValueIsZero(RHS, Mask))
3634 case Instruction::Or:
3635 case Instruction::Xor:
3636 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3637 if ((Mask->getValue().countLeadingZeros() +
3638 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3639 && And(N, Mask)->isZero())
3646 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
3648 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
3649 return InsertNewInstBefore(New, I);
3652 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
3653 Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
3654 ICmpInst *LHS, ICmpInst *RHS) {
3656 ConstantInt *LHSCst, *RHSCst;
3657 ICmpInst::Predicate LHSCC, RHSCC;
3659 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
3660 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val), m_ConstantInt(LHSCst))) ||
3661 !match(RHS, m_ICmp(RHSCC, m_Value(Val2), m_ConstantInt(RHSCst))))
3664 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
3665 // where C is a power of 2
3666 if (LHSCst == RHSCst && LHSCC == RHSCC && LHSCC == ICmpInst::ICMP_ULT &&
3667 LHSCst->getValue().isPowerOf2()) {
3668 Instruction *NewOr = BinaryOperator::CreateOr(Val, Val2);
3669 InsertNewInstBefore(NewOr, I);
3670 return new ICmpInst(LHSCC, NewOr, LHSCst);
3673 // From here on, we only handle:
3674 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
3675 if (Val != Val2) return 0;
3677 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
3678 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
3679 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
3680 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
3681 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
3684 // We can't fold (ugt x, C) & (sgt x, C2).
3685 if (!PredicatesFoldable(LHSCC, RHSCC))
3688 // Ensure that the larger constant is on the RHS.
3690 if (ICmpInst::isSignedPredicate(LHSCC) ||
3691 (ICmpInst::isEquality(LHSCC) &&
3692 ICmpInst::isSignedPredicate(RHSCC)))
3693 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3695 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3698 std::swap(LHS, RHS);
3699 std::swap(LHSCst, RHSCst);
3700 std::swap(LHSCC, RHSCC);
3703 // At this point, we know we have have two icmp instructions
3704 // comparing a value against two constants and and'ing the result
3705 // together. Because of the above check, we know that we only have
3706 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3707 // (from the FoldICmpLogical check above), that the two constants
3708 // are not equal and that the larger constant is on the RHS
3709 assert(LHSCst != RHSCst && "Compares not folded above?");
3712 default: assert(0 && "Unknown integer condition code!");
3713 case ICmpInst::ICMP_EQ:
3715 default: assert(0 && "Unknown integer condition code!");
3716 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3717 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3718 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3719 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3720 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3721 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3722 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3723 return ReplaceInstUsesWith(I, LHS);
3725 case ICmpInst::ICMP_NE:
3727 default: assert(0 && "Unknown integer condition code!");
3728 case ICmpInst::ICMP_ULT:
3729 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3730 return new ICmpInst(ICmpInst::ICMP_ULT, Val, LHSCst);
3731 break; // (X != 13 & X u< 15) -> no change
3732 case ICmpInst::ICMP_SLT:
3733 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3734 return new ICmpInst(ICmpInst::ICMP_SLT, Val, LHSCst);
3735 break; // (X != 13 & X s< 15) -> no change
3736 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3737 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3738 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3739 return ReplaceInstUsesWith(I, RHS);
3740 case ICmpInst::ICMP_NE:
3741 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3742 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3743 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
3744 Val->getName()+".off");
3745 InsertNewInstBefore(Add, I);
3746 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3747 ConstantInt::get(Add->getType(), 1));
3749 break; // (X != 13 & X != 15) -> no change
3752 case ICmpInst::ICMP_ULT:
3754 default: assert(0 && "Unknown integer condition code!");
3755 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3756 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3757 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3758 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3760 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3761 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3762 return ReplaceInstUsesWith(I, LHS);
3763 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3767 case ICmpInst::ICMP_SLT:
3769 default: assert(0 && "Unknown integer condition code!");
3770 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3771 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3772 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3773 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3775 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3776 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3777 return ReplaceInstUsesWith(I, LHS);
3778 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3782 case ICmpInst::ICMP_UGT:
3784 default: assert(0 && "Unknown integer condition code!");
3785 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3786 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3787 return ReplaceInstUsesWith(I, RHS);
3788 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3790 case ICmpInst::ICMP_NE:
3791 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3792 return new ICmpInst(LHSCC, Val, RHSCst);
3793 break; // (X u> 13 & X != 15) -> no change
3794 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
3795 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true, I);
3796 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3800 case ICmpInst::ICMP_SGT:
3802 default: assert(0 && "Unknown integer condition code!");
3803 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3804 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3805 return ReplaceInstUsesWith(I, RHS);
3806 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3808 case ICmpInst::ICMP_NE:
3809 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3810 return new ICmpInst(LHSCC, Val, RHSCst);
3811 break; // (X s> 13 & X != 15) -> no change
3812 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
3813 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true, I);
3814 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3824 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3825 bool Changed = SimplifyCommutative(I);
3826 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3828 if (isa<UndefValue>(Op1)) // X & undef -> 0
3829 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3833 return ReplaceInstUsesWith(I, Op1);
3835 // See if we can simplify any instructions used by the instruction whose sole
3836 // purpose is to compute bits we don't care about.
3837 if (!isa<VectorType>(I.getType())) {
3838 if (SimplifyDemandedInstructionBits(I))
3841 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3842 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3843 return ReplaceInstUsesWith(I, I.getOperand(0));
3844 } else if (isa<ConstantAggregateZero>(Op1)) {
3845 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3849 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3850 const APInt& AndRHSMask = AndRHS->getValue();
3851 APInt NotAndRHS(~AndRHSMask);
3853 // Optimize a variety of ((val OP C1) & C2) combinations...
3854 if (isa<BinaryOperator>(Op0)) {
3855 Instruction *Op0I = cast<Instruction>(Op0);
3856 Value *Op0LHS = Op0I->getOperand(0);
3857 Value *Op0RHS = Op0I->getOperand(1);
3858 switch (Op0I->getOpcode()) {
3859 case Instruction::Xor:
3860 case Instruction::Or:
3861 // If the mask is only needed on one incoming arm, push it up.
3862 if (Op0I->hasOneUse()) {
3863 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3864 // Not masking anything out for the LHS, move to RHS.
3865 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
3866 Op0RHS->getName()+".masked");
3867 InsertNewInstBefore(NewRHS, I);
3868 return BinaryOperator::Create(
3869 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3871 if (!isa<Constant>(Op0RHS) &&
3872 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3873 // Not masking anything out for the RHS, move to LHS.
3874 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
3875 Op0LHS->getName()+".masked");
3876 InsertNewInstBefore(NewLHS, I);
3877 return BinaryOperator::Create(
3878 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3883 case Instruction::Add:
3884 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3885 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3886 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3887 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3888 return BinaryOperator::CreateAnd(V, AndRHS);
3889 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3890 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
3893 case Instruction::Sub:
3894 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3895 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3896 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3897 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3898 return BinaryOperator::CreateAnd(V, AndRHS);
3900 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
3901 // has 1's for all bits that the subtraction with A might affect.
3902 if (Op0I->hasOneUse()) {
3903 uint32_t BitWidth = AndRHSMask.getBitWidth();
3904 uint32_t Zeros = AndRHSMask.countLeadingZeros();
3905 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
3907 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
3908 if (!(A && A->isZero()) && // avoid infinite recursion.
3909 MaskedValueIsZero(Op0LHS, Mask)) {
3910 Instruction *NewNeg = BinaryOperator::CreateNeg(Op0RHS);
3911 InsertNewInstBefore(NewNeg, I);
3912 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
3917 case Instruction::Shl:
3918 case Instruction::LShr:
3919 // (1 << x) & 1 --> zext(x == 0)
3920 // (1 >> x) & 1 --> zext(x == 0)
3921 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
3922 Instruction *NewICmp = new ICmpInst(ICmpInst::ICMP_EQ, Op0RHS,
3923 Constant::getNullValue(I.getType()));
3924 InsertNewInstBefore(NewICmp, I);
3925 return new ZExtInst(NewICmp, I.getType());
3930 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3931 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3933 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3934 // If this is an integer truncation or change from signed-to-unsigned, and
3935 // if the source is an and/or with immediate, transform it. This
3936 // frequently occurs for bitfield accesses.
3937 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3938 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3939 CastOp->getNumOperands() == 2)
3940 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
3941 if (CastOp->getOpcode() == Instruction::And) {
3942 // Change: and (cast (and X, C1) to T), C2
3943 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3944 // This will fold the two constants together, which may allow
3945 // other simplifications.
3946 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
3947 CastOp->getOperand(0), I.getType(),
3948 CastOp->getName()+".shrunk");
3949 NewCast = InsertNewInstBefore(NewCast, I);
3950 // trunc_or_bitcast(C1)&C2
3951 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3952 C3 = ConstantExpr::getAnd(C3, AndRHS);
3953 return BinaryOperator::CreateAnd(NewCast, C3);
3954 } else if (CastOp->getOpcode() == Instruction::Or) {
3955 // Change: and (cast (or X, C1) to T), C2
3956 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3957 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3958 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3959 return ReplaceInstUsesWith(I, AndRHS);
3965 // Try to fold constant and into select arguments.
3966 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3967 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3969 if (isa<PHINode>(Op0))
3970 if (Instruction *NV = FoldOpIntoPhi(I))
3974 Value *Op0NotVal = dyn_castNotVal(Op0);
3975 Value *Op1NotVal = dyn_castNotVal(Op1);
3977 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3978 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3980 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3981 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3982 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
3983 I.getName()+".demorgan");
3984 InsertNewInstBefore(Or, I);
3985 return BinaryOperator::CreateNot(Or);
3989 Value *A = 0, *B = 0, *C = 0, *D = 0;
3990 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3991 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3992 return ReplaceInstUsesWith(I, Op1);
3994 // (A|B) & ~(A&B) -> A^B
3995 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3996 if ((A == C && B == D) || (A == D && B == C))
3997 return BinaryOperator::CreateXor(A, B);
4001 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
4002 if (A == Op0 || B == Op0) // A & (A | ?) --> A
4003 return ReplaceInstUsesWith(I, Op0);
4005 // ~(A&B) & (A|B) -> A^B
4006 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
4007 if ((A == C && B == D) || (A == D && B == C))
4008 return BinaryOperator::CreateXor(A, B);
4012 if (Op0->hasOneUse() &&
4013 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4014 if (A == Op1) { // (A^B)&A -> A&(A^B)
4015 I.swapOperands(); // Simplify below
4016 std::swap(Op0, Op1);
4017 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
4018 cast<BinaryOperator>(Op0)->swapOperands();
4019 I.swapOperands(); // Simplify below
4020 std::swap(Op0, Op1);
4024 if (Op1->hasOneUse() &&
4025 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
4026 if (B == Op0) { // B&(A^B) -> B&(B^A)
4027 cast<BinaryOperator>(Op1)->swapOperands();
4030 if (A == Op0) { // A&(A^B) -> A & ~B
4031 Instruction *NotB = BinaryOperator::CreateNot(B, "tmp");
4032 InsertNewInstBefore(NotB, I);
4033 return BinaryOperator::CreateAnd(A, NotB);
4037 // (A&((~A)|B)) -> A&B
4038 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
4039 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
4040 return BinaryOperator::CreateAnd(A, Op1);
4041 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
4042 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
4043 return BinaryOperator::CreateAnd(A, Op0);
4046 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
4047 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
4048 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4051 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
4052 if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
4056 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4057 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4058 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4059 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4060 const Type *SrcTy = Op0C->getOperand(0)->getType();
4061 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4062 // Only do this if the casts both really cause code to be generated.
4063 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4065 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4067 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
4068 Op1C->getOperand(0),
4070 InsertNewInstBefore(NewOp, I);
4071 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4075 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4076 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4077 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4078 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4079 SI0->getOperand(1) == SI1->getOperand(1) &&
4080 (SI0->hasOneUse() || SI1->hasOneUse())) {
4081 Instruction *NewOp =
4082 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
4084 SI0->getName()), I);
4085 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4086 SI1->getOperand(1));
4090 // If and'ing two fcmp, try combine them into one.
4091 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4092 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4093 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
4094 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
4095 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
4096 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4097 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4098 // If either of the constants are nans, then the whole thing returns
4100 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4101 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4102 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
4103 RHS->getOperand(0));
4106 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4107 FCmpInst::Predicate Op0CC, Op1CC;
4108 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS), m_Value(Op0RHS))) &&
4109 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS), m_Value(Op1RHS)))) {
4110 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4111 // Swap RHS operands to match LHS.
4112 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4113 std::swap(Op1LHS, Op1RHS);
4115 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4116 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
4118 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4119 else if (Op0CC == FCmpInst::FCMP_FALSE ||
4120 Op1CC == FCmpInst::FCMP_FALSE)
4121 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4122 else if (Op0CC == FCmpInst::FCMP_TRUE)
4123 return ReplaceInstUsesWith(I, Op1);
4124 else if (Op1CC == FCmpInst::FCMP_TRUE)
4125 return ReplaceInstUsesWith(I, Op0);
4128 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4129 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4131 std::swap(Op0, Op1);
4132 std::swap(Op0Pred, Op1Pred);
4133 std::swap(Op0Ordered, Op1Ordered);
4136 // uno && ueq -> uno && (uno || eq) -> ueq
4137 // ord && olt -> ord && (ord && lt) -> olt
4138 if (Op0Ordered == Op1Ordered)
4139 return ReplaceInstUsesWith(I, Op1);
4140 // uno && oeq -> uno && (ord && eq) -> false
4141 // uno && ord -> false
4143 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4144 // ord && ueq -> ord && (uno || eq) -> oeq
4145 return cast<Instruction>(getFCmpValue(true, Op1Pred,
4154 return Changed ? &I : 0;
4157 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
4158 /// capable of providing pieces of a bswap. The subexpression provides pieces
4159 /// of a bswap if it is proven that each of the non-zero bytes in the output of
4160 /// the expression came from the corresponding "byte swapped" byte in some other
4161 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
4162 /// we know that the expression deposits the low byte of %X into the high byte
4163 /// of the bswap result and that all other bytes are zero. This expression is
4164 /// accepted, the high byte of ByteValues is set to X to indicate a correct
4167 /// This function returns true if the match was unsuccessful and false if so.
4168 /// On entry to the function the "OverallLeftShift" is a signed integer value
4169 /// indicating the number of bytes that the subexpression is later shifted. For
4170 /// example, if the expression is later right shifted by 16 bits, the
4171 /// OverallLeftShift value would be -2 on entry. This is used to specify which
4172 /// byte of ByteValues is actually being set.
4174 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
4175 /// byte is masked to zero by a user. For example, in (X & 255), X will be
4176 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
4177 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
4178 /// always in the local (OverallLeftShift) coordinate space.
4180 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
4181 SmallVector<Value*, 8> &ByteValues) {
4182 if (Instruction *I = dyn_cast<Instruction>(V)) {
4183 // If this is an or instruction, it may be an inner node of the bswap.
4184 if (I->getOpcode() == Instruction::Or) {
4185 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4187 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
4191 // If this is a logical shift by a constant multiple of 8, recurse with
4192 // OverallLeftShift and ByteMask adjusted.
4193 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
4195 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
4196 // Ensure the shift amount is defined and of a byte value.
4197 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
4200 unsigned ByteShift = ShAmt >> 3;
4201 if (I->getOpcode() == Instruction::Shl) {
4202 // X << 2 -> collect(X, +2)
4203 OverallLeftShift += ByteShift;
4204 ByteMask >>= ByteShift;
4206 // X >>u 2 -> collect(X, -2)
4207 OverallLeftShift -= ByteShift;
4208 ByteMask <<= ByteShift;
4209 ByteMask &= (~0U >> (32-ByteValues.size()));
4212 if (OverallLeftShift >= (int)ByteValues.size()) return true;
4213 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
4215 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4219 // If this is a logical 'and' with a mask that clears bytes, clear the
4220 // corresponding bytes in ByteMask.
4221 if (I->getOpcode() == Instruction::And &&
4222 isa<ConstantInt>(I->getOperand(1))) {
4223 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
4224 unsigned NumBytes = ByteValues.size();
4225 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
4226 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
4228 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
4229 // If this byte is masked out by a later operation, we don't care what
4231 if ((ByteMask & (1 << i)) == 0)
4234 // If the AndMask is all zeros for this byte, clear the bit.
4235 APInt MaskB = AndMask & Byte;
4237 ByteMask &= ~(1U << i);
4241 // If the AndMask is not all ones for this byte, it's not a bytezap.
4245 // Otherwise, this byte is kept.
4248 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4253 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
4254 // the input value to the bswap. Some observations: 1) if more than one byte
4255 // is demanded from this input, then it could not be successfully assembled
4256 // into a byteswap. At least one of the two bytes would not be aligned with
4257 // their ultimate destination.
4258 if (!isPowerOf2_32(ByteMask)) return true;
4259 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
4261 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
4262 // is demanded, it needs to go into byte 0 of the result. This means that the
4263 // byte needs to be shifted until it lands in the right byte bucket. The
4264 // shift amount depends on the position: if the byte is coming from the high
4265 // part of the value (e.g. byte 3) then it must be shifted right. If from the
4266 // low part, it must be shifted left.
4267 unsigned DestByteNo = InputByteNo + OverallLeftShift;
4268 if (InputByteNo < ByteValues.size()/2) {
4269 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4272 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4276 // If the destination byte value is already defined, the values are or'd
4277 // together, which isn't a bswap (unless it's an or of the same bits).
4278 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
4280 ByteValues[DestByteNo] = V;
4284 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4285 /// If so, insert the new bswap intrinsic and return it.
4286 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4287 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4288 if (!ITy || ITy->getBitWidth() % 16 ||
4289 // ByteMask only allows up to 32-byte values.
4290 ITy->getBitWidth() > 32*8)
4291 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4293 /// ByteValues - For each byte of the result, we keep track of which value
4294 /// defines each byte.
4295 SmallVector<Value*, 8> ByteValues;
4296 ByteValues.resize(ITy->getBitWidth()/8);
4298 // Try to find all the pieces corresponding to the bswap.
4299 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
4300 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
4303 // Check to see if all of the bytes come from the same value.
4304 Value *V = ByteValues[0];
4305 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4307 // Check to make sure that all of the bytes come from the same value.
4308 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4309 if (ByteValues[i] != V)
4311 const Type *Tys[] = { ITy };
4312 Module *M = I.getParent()->getParent()->getParent();
4313 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4314 return CallInst::Create(F, V);
4317 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
4318 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
4319 /// we can simplify this expression to "cond ? C : D or B".
4320 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
4321 Value *C, Value *D) {
4322 // If A is not a select of -1/0, this cannot match.
4324 if (!match(A, m_SelectCst<-1, 0>(m_Value(Cond))))
4327 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
4328 if (match(D, m_SelectCst<0, -1>(m_Specific(Cond))))
4329 return SelectInst::Create(Cond, C, B);
4330 if (match(D, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
4331 return SelectInst::Create(Cond, C, B);
4332 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
4333 if (match(B, m_SelectCst<0, -1>(m_Specific(Cond))))
4334 return SelectInst::Create(Cond, C, D);
4335 if (match(B, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
4336 return SelectInst::Create(Cond, C, D);
4340 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
4341 Instruction *InstCombiner::FoldOrOfICmps(Instruction &I,
4342 ICmpInst *LHS, ICmpInst *RHS) {
4344 ConstantInt *LHSCst, *RHSCst;
4345 ICmpInst::Predicate LHSCC, RHSCC;
4347 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
4348 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val), m_ConstantInt(LHSCst))) ||
4349 !match(RHS, m_ICmp(RHSCC, m_Value(Val2), m_ConstantInt(RHSCst))))
4352 // From here on, we only handle:
4353 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
4354 if (Val != Val2) return 0;
4356 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
4357 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
4358 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
4359 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
4360 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
4363 // We can't fold (ugt x, C) | (sgt x, C2).
4364 if (!PredicatesFoldable(LHSCC, RHSCC))
4367 // Ensure that the larger constant is on the RHS.
4369 if (ICmpInst::isSignedPredicate(LHSCC) ||
4370 (ICmpInst::isEquality(LHSCC) &&
4371 ICmpInst::isSignedPredicate(RHSCC)))
4372 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4374 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4377 std::swap(LHS, RHS);
4378 std::swap(LHSCst, RHSCst);
4379 std::swap(LHSCC, RHSCC);
4382 // At this point, we know we have have two icmp instructions
4383 // comparing a value against two constants and or'ing the result
4384 // together. Because of the above check, we know that we only have
4385 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4386 // FoldICmpLogical check above), that the two constants are not
4388 assert(LHSCst != RHSCst && "Compares not folded above?");
4391 default: assert(0 && "Unknown integer condition code!");
4392 case ICmpInst::ICMP_EQ:
4394 default: assert(0 && "Unknown integer condition code!");
4395 case ICmpInst::ICMP_EQ:
4396 if (LHSCst == SubOne(RHSCst)) { // (X == 13 | X == 14) -> X-13 <u 2
4397 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4398 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
4399 Val->getName()+".off");
4400 InsertNewInstBefore(Add, I);
4401 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4402 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4404 break; // (X == 13 | X == 15) -> no change
4405 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4406 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4408 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4409 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4410 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4411 return ReplaceInstUsesWith(I, RHS);
4414 case ICmpInst::ICMP_NE:
4416 default: assert(0 && "Unknown integer condition code!");
4417 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4418 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4419 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4420 return ReplaceInstUsesWith(I, LHS);
4421 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4422 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4423 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4424 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4427 case ICmpInst::ICMP_ULT:
4429 default: assert(0 && "Unknown integer condition code!");
4430 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4432 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
4433 // If RHSCst is [us]MAXINT, it is always false. Not handling
4434 // this can cause overflow.
4435 if (RHSCst->isMaxValue(false))
4436 return ReplaceInstUsesWith(I, LHS);
4437 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false, false, I);
4438 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4440 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4441 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4442 return ReplaceInstUsesWith(I, RHS);
4443 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4447 case ICmpInst::ICMP_SLT:
4449 default: assert(0 && "Unknown integer condition code!");
4450 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4452 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
4453 // If RHSCst is [us]MAXINT, it is always false. Not handling
4454 // this can cause overflow.
4455 if (RHSCst->isMaxValue(true))
4456 return ReplaceInstUsesWith(I, LHS);
4457 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true, false, I);
4458 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4460 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4461 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4462 return ReplaceInstUsesWith(I, RHS);
4463 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4467 case ICmpInst::ICMP_UGT:
4469 default: assert(0 && "Unknown integer condition code!");
4470 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4471 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4472 return ReplaceInstUsesWith(I, LHS);
4473 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4475 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4476 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4477 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4478 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4482 case ICmpInst::ICMP_SGT:
4484 default: assert(0 && "Unknown integer condition code!");
4485 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4486 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4487 return ReplaceInstUsesWith(I, LHS);
4488 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4490 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4491 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4492 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4493 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4501 /// FoldOrWithConstants - This helper function folds:
4503 /// ((A | B) & C1) | (B & C2)
4509 /// when the XOR of the two constants is "all ones" (-1).
4510 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
4511 Value *A, Value *B, Value *C) {
4512 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
4516 ConstantInt *CI2 = 0;
4517 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return 0;
4519 APInt Xor = CI1->getValue() ^ CI2->getValue();
4520 if (!Xor.isAllOnesValue()) return 0;
4522 if (V1 == A || V1 == B) {
4523 Instruction *NewOp =
4524 InsertNewInstBefore(BinaryOperator::CreateAnd((V1 == A) ? B : A, CI1), I);
4525 return BinaryOperator::CreateOr(NewOp, V1);
4531 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4532 bool Changed = SimplifyCommutative(I);
4533 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4535 if (isa<UndefValue>(Op1)) // X | undef -> -1
4536 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4540 return ReplaceInstUsesWith(I, Op0);
4542 // See if we can simplify any instructions used by the instruction whose sole
4543 // purpose is to compute bits we don't care about.
4544 if (!isa<VectorType>(I.getType())) {
4545 if (SimplifyDemandedInstructionBits(I))
4547 } else if (isa<ConstantAggregateZero>(Op1)) {
4548 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4549 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4550 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4551 return ReplaceInstUsesWith(I, I.getOperand(1));
4557 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4558 ConstantInt *C1 = 0; Value *X = 0;
4559 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4560 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4561 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4562 InsertNewInstBefore(Or, I);
4564 return BinaryOperator::CreateAnd(Or,
4565 ConstantInt::get(RHS->getValue() | C1->getValue()));
4568 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4569 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4570 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4571 InsertNewInstBefore(Or, I);
4573 return BinaryOperator::CreateXor(Or,
4574 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
4577 // Try to fold constant and into select arguments.
4578 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4579 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4581 if (isa<PHINode>(Op0))
4582 if (Instruction *NV = FoldOpIntoPhi(I))
4586 Value *A = 0, *B = 0;
4587 ConstantInt *C1 = 0, *C2 = 0;
4589 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4590 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4591 return ReplaceInstUsesWith(I, Op1);
4592 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4593 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4594 return ReplaceInstUsesWith(I, Op0);
4596 // (A | B) | C and A | (B | C) -> bswap if possible.
4597 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4598 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4599 match(Op1, m_Or(m_Value(), m_Value())) ||
4600 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4601 match(Op1, m_Shift(m_Value(), m_Value())))) {
4602 if (Instruction *BSwap = MatchBSwap(I))
4606 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4607 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4608 MaskedValueIsZero(Op1, C1->getValue())) {
4609 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
4610 InsertNewInstBefore(NOr, I);
4612 return BinaryOperator::CreateXor(NOr, C1);
4615 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4616 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4617 MaskedValueIsZero(Op0, C1->getValue())) {
4618 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
4619 InsertNewInstBefore(NOr, I);
4621 return BinaryOperator::CreateXor(NOr, C1);
4625 Value *C = 0, *D = 0;
4626 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4627 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4628 Value *V1 = 0, *V2 = 0, *V3 = 0;
4629 C1 = dyn_cast<ConstantInt>(C);
4630 C2 = dyn_cast<ConstantInt>(D);
4631 if (C1 && C2) { // (A & C1)|(B & C2)
4632 // If we have: ((V + N) & C1) | (V & C2)
4633 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4634 // replace with V+N.
4635 if (C1->getValue() == ~C2->getValue()) {
4636 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4637 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4638 // Add commutes, try both ways.
4639 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4640 return ReplaceInstUsesWith(I, A);
4641 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4642 return ReplaceInstUsesWith(I, A);
4644 // Or commutes, try both ways.
4645 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4646 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4647 // Add commutes, try both ways.
4648 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4649 return ReplaceInstUsesWith(I, B);
4650 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4651 return ReplaceInstUsesWith(I, B);
4654 V1 = 0; V2 = 0; V3 = 0;
4657 // Check to see if we have any common things being and'ed. If so, find the
4658 // terms for V1 & (V2|V3).
4659 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4660 if (A == B) // (A & C)|(A & D) == A & (C|D)
4661 V1 = A, V2 = C, V3 = D;
4662 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4663 V1 = A, V2 = B, V3 = C;
4664 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4665 V1 = C, V2 = A, V3 = D;
4666 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4667 V1 = C, V2 = A, V3 = B;
4671 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4672 return BinaryOperator::CreateAnd(V1, Or);
4676 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
4677 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D))
4679 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C))
4681 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D))
4683 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C))
4686 // ((A&~B)|(~A&B)) -> A^B
4687 if ((match(C, m_Not(m_Specific(D))) &&
4688 match(B, m_Not(m_Specific(A)))))
4689 return BinaryOperator::CreateXor(A, D);
4690 // ((~B&A)|(~A&B)) -> A^B
4691 if ((match(A, m_Not(m_Specific(D))) &&
4692 match(B, m_Not(m_Specific(C)))))
4693 return BinaryOperator::CreateXor(C, D);
4694 // ((A&~B)|(B&~A)) -> A^B
4695 if ((match(C, m_Not(m_Specific(B))) &&
4696 match(D, m_Not(m_Specific(A)))))
4697 return BinaryOperator::CreateXor(A, B);
4698 // ((~B&A)|(B&~A)) -> A^B
4699 if ((match(A, m_Not(m_Specific(B))) &&
4700 match(D, m_Not(m_Specific(C)))))
4701 return BinaryOperator::CreateXor(C, B);
4704 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4705 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4706 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4707 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4708 SI0->getOperand(1) == SI1->getOperand(1) &&
4709 (SI0->hasOneUse() || SI1->hasOneUse())) {
4710 Instruction *NewOp =
4711 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4713 SI0->getName()), I);
4714 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4715 SI1->getOperand(1));
4719 // ((A|B)&1)|(B&-2) -> (A&1) | B
4720 if (match(Op0, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
4721 match(Op0, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
4722 Instruction *Ret = FoldOrWithConstants(I, Op1, A, B, C);
4723 if (Ret) return Ret;
4725 // (B&-2)|((A|B)&1) -> (A&1) | B
4726 if (match(Op1, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
4727 match(Op1, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
4728 Instruction *Ret = FoldOrWithConstants(I, Op0, A, B, C);
4729 if (Ret) return Ret;
4732 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4733 if (A == Op1) // ~A | A == -1
4734 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4738 // Note, A is still live here!
4739 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4741 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4743 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4744 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4745 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4746 I.getName()+".demorgan"), I);
4747 return BinaryOperator::CreateNot(And);
4751 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4752 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4753 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4756 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
4757 if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
4761 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4762 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4763 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4764 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4765 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4766 !isa<ICmpInst>(Op1C->getOperand(0))) {
4767 const Type *SrcTy = Op0C->getOperand(0)->getType();
4768 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4769 // Only do this if the casts both really cause code to be
4771 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4773 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4775 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4776 Op1C->getOperand(0),
4778 InsertNewInstBefore(NewOp, I);
4779 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4786 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4787 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4788 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4789 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4790 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4791 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
4792 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4793 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4794 // If either of the constants are nans, then the whole thing returns
4796 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4797 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4799 // Otherwise, no need to compare the two constants, compare the
4801 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4802 RHS->getOperand(0));
4805 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4806 FCmpInst::Predicate Op0CC, Op1CC;
4807 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS), m_Value(Op0RHS))) &&
4808 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS), m_Value(Op1RHS)))) {
4809 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4810 // Swap RHS operands to match LHS.
4811 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4812 std::swap(Op1LHS, Op1RHS);
4814 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4815 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
4817 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4818 else if (Op0CC == FCmpInst::FCMP_TRUE ||
4819 Op1CC == FCmpInst::FCMP_TRUE)
4820 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4821 else if (Op0CC == FCmpInst::FCMP_FALSE)
4822 return ReplaceInstUsesWith(I, Op1);
4823 else if (Op1CC == FCmpInst::FCMP_FALSE)
4824 return ReplaceInstUsesWith(I, Op0);
4827 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4828 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4829 if (Op0Ordered == Op1Ordered) {
4830 // If both are ordered or unordered, return a new fcmp with
4831 // or'ed predicates.
4832 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred,
4834 if (Instruction *I = dyn_cast<Instruction>(RV))
4836 // Otherwise, it's a constant boolean value...
4837 return ReplaceInstUsesWith(I, RV);
4845 return Changed ? &I : 0;
4850 // XorSelf - Implements: X ^ X --> 0
4853 XorSelf(Value *rhs) : RHS(rhs) {}
4854 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4855 Instruction *apply(BinaryOperator &Xor) const {
4862 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4863 bool Changed = SimplifyCommutative(I);
4864 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4866 if (isa<UndefValue>(Op1)) {
4867 if (isa<UndefValue>(Op0))
4868 // Handle undef ^ undef -> 0 special case. This is a common
4870 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4871 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4874 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4875 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4876 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4877 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4880 // See if we can simplify any instructions used by the instruction whose sole
4881 // purpose is to compute bits we don't care about.
4882 if (!isa<VectorType>(I.getType())) {
4883 if (SimplifyDemandedInstructionBits(I))
4885 } else if (isa<ConstantAggregateZero>(Op1)) {
4886 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4889 // Is this a ~ operation?
4890 if (Value *NotOp = dyn_castNotVal(&I)) {
4891 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4892 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4893 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4894 if (Op0I->getOpcode() == Instruction::And ||
4895 Op0I->getOpcode() == Instruction::Or) {
4896 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4897 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4899 BinaryOperator::CreateNot(Op0I->getOperand(1),
4900 Op0I->getOperand(1)->getName()+".not");
4901 InsertNewInstBefore(NotY, I);
4902 if (Op0I->getOpcode() == Instruction::And)
4903 return BinaryOperator::CreateOr(Op0NotVal, NotY);
4905 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
4912 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4913 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4914 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4915 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4916 return new ICmpInst(ICI->getInversePredicate(),
4917 ICI->getOperand(0), ICI->getOperand(1));
4919 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4920 return new FCmpInst(FCI->getInversePredicate(),
4921 FCI->getOperand(0), FCI->getOperand(1));
4924 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
4925 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4926 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
4927 if (CI->hasOneUse() && Op0C->hasOneUse()) {
4928 Instruction::CastOps Opcode = Op0C->getOpcode();
4929 if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
4930 if (RHS == ConstantExpr::getCast(Opcode, ConstantInt::getTrue(),
4931 Op0C->getDestTy())) {
4932 Instruction *NewCI = InsertNewInstBefore(CmpInst::Create(
4933 CI->getOpcode(), CI->getInversePredicate(),
4934 CI->getOperand(0), CI->getOperand(1)), I);
4935 NewCI->takeName(CI);
4936 return CastInst::Create(Opcode, NewCI, Op0C->getType());
4943 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4944 // ~(c-X) == X-c-1 == X+(-c-1)
4945 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4946 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4947 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4948 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4949 ConstantInt::get(I.getType(), 1));
4950 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
4953 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
4954 if (Op0I->getOpcode() == Instruction::Add) {
4955 // ~(X-c) --> (-c-1)-X
4956 if (RHS->isAllOnesValue()) {
4957 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4958 return BinaryOperator::CreateSub(
4959 ConstantExpr::getSub(NegOp0CI,
4960 ConstantInt::get(I.getType(), 1)),
4961 Op0I->getOperand(0));
4962 } else if (RHS->getValue().isSignBit()) {
4963 // (X + C) ^ signbit -> (X + C + signbit)
4964 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4965 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
4968 } else if (Op0I->getOpcode() == Instruction::Or) {
4969 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4970 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4971 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4972 // Anything in both C1 and C2 is known to be zero, remove it from
4974 Constant *CommonBits = And(Op0CI, RHS);
4975 NewRHS = ConstantExpr::getAnd(NewRHS,
4976 ConstantExpr::getNot(CommonBits));
4977 AddToWorkList(Op0I);
4978 I.setOperand(0, Op0I->getOperand(0));
4979 I.setOperand(1, NewRHS);
4986 // Try to fold constant and into select arguments.
4987 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4988 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4990 if (isa<PHINode>(Op0))
4991 if (Instruction *NV = FoldOpIntoPhi(I))
4995 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4997 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4999 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
5001 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5004 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
5007 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
5008 if (A == Op0) { // B^(B|A) == (A|B)^B
5009 Op1I->swapOperands();
5011 std::swap(Op0, Op1);
5012 } else if (B == Op0) { // B^(A|B) == (A|B)^B
5013 I.swapOperands(); // Simplified below.
5014 std::swap(Op0, Op1);
5016 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)))) {
5017 return ReplaceInstUsesWith(I, B); // A^(A^B) == B
5018 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)))) {
5019 return ReplaceInstUsesWith(I, A); // A^(B^A) == B
5020 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
5021 if (A == Op0) { // A^(A&B) -> A^(B&A)
5022 Op1I->swapOperands();
5025 if (B == Op0) { // A^(B&A) -> (B&A)^A
5026 I.swapOperands(); // Simplified below.
5027 std::swap(Op0, Op1);
5032 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
5035 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
5036 if (A == Op1) // (B|A)^B == (A|B)^B
5038 if (B == Op1) { // (A|B)^B == A & ~B
5040 InsertNewInstBefore(BinaryOperator::CreateNot(Op1, "tmp"), I);
5041 return BinaryOperator::CreateAnd(A, NotB);
5043 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)))) {
5044 return ReplaceInstUsesWith(I, B); // (A^B)^A == B
5045 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)))) {
5046 return ReplaceInstUsesWith(I, A); // (B^A)^A == B
5047 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
5048 if (A == Op1) // (A&B)^A -> (B&A)^A
5050 if (B == Op1 && // (B&A)^A == ~B & A
5051 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
5053 InsertNewInstBefore(BinaryOperator::CreateNot(A, "tmp"), I);
5054 return BinaryOperator::CreateAnd(N, Op1);
5059 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
5060 if (Op0I && Op1I && Op0I->isShift() &&
5061 Op0I->getOpcode() == Op1I->getOpcode() &&
5062 Op0I->getOperand(1) == Op1I->getOperand(1) &&
5063 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
5064 Instruction *NewOp =
5065 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
5066 Op1I->getOperand(0),
5067 Op0I->getName()), I);
5068 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
5069 Op1I->getOperand(1));
5073 Value *A, *B, *C, *D;
5074 // (A & B)^(A | B) -> A ^ B
5075 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5076 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
5077 if ((A == C && B == D) || (A == D && B == C))
5078 return BinaryOperator::CreateXor(A, B);
5080 // (A | B)^(A & B) -> A ^ B
5081 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5082 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5083 if ((A == C && B == D) || (A == D && B == C))
5084 return BinaryOperator::CreateXor(A, B);
5088 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
5089 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5090 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5091 // (X & Y)^(X & Y) -> (Y^Z) & X
5092 Value *X = 0, *Y = 0, *Z = 0;
5094 X = A, Y = B, Z = D;
5096 X = A, Y = B, Z = C;
5098 X = B, Y = A, Z = D;
5100 X = B, Y = A, Z = C;
5103 Instruction *NewOp =
5104 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
5105 return BinaryOperator::CreateAnd(NewOp, X);
5110 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
5111 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
5112 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
5115 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
5116 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5117 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5118 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
5119 const Type *SrcTy = Op0C->getOperand(0)->getType();
5120 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
5121 // Only do this if the casts both really cause code to be generated.
5122 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5124 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5126 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
5127 Op1C->getOperand(0),
5129 InsertNewInstBefore(NewOp, I);
5130 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5135 return Changed ? &I : 0;
5138 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
5139 /// overflowed for this type.
5140 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
5141 ConstantInt *In2, bool IsSigned = false) {
5142 Result = cast<ConstantInt>(Add(In1, In2));
5145 if (In2->getValue().isNegative())
5146 return Result->getValue().sgt(In1->getValue());
5148 return Result->getValue().slt(In1->getValue());
5150 return Result->getValue().ult(In1->getValue());
5153 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
5154 /// overflowed for this type.
5155 static bool SubWithOverflow(ConstantInt *&Result, ConstantInt *In1,
5156 ConstantInt *In2, bool IsSigned = false) {
5157 Result = cast<ConstantInt>(Subtract(In1, In2));
5160 if (In2->getValue().isNegative())
5161 return Result->getValue().slt(In1->getValue());
5163 return Result->getValue().sgt(In1->getValue());
5165 return Result->getValue().ugt(In1->getValue());
5168 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
5169 /// code necessary to compute the offset from the base pointer (without adding
5170 /// in the base pointer). Return the result as a signed integer of intptr size.
5171 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
5172 TargetData &TD = IC.getTargetData();
5173 gep_type_iterator GTI = gep_type_begin(GEP);
5174 const Type *IntPtrTy = TD.getIntPtrType();
5175 Value *Result = Constant::getNullValue(IntPtrTy);
5177 // Build a mask for high order bits.
5178 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5179 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5181 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
5184 uint64_t Size = TD.getTypePaddedSize(GTI.getIndexedType()) & PtrSizeMask;
5185 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
5186 if (OpC->isZero()) continue;
5188 // Handle a struct index, which adds its field offset to the pointer.
5189 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5190 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
5192 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
5193 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
5195 Result = IC.InsertNewInstBefore(
5196 BinaryOperator::CreateAdd(Result,
5197 ConstantInt::get(IntPtrTy, Size),
5198 GEP->getName()+".offs"), I);
5202 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5203 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
5204 Scale = ConstantExpr::getMul(OC, Scale);
5205 if (Constant *RC = dyn_cast<Constant>(Result))
5206 Result = ConstantExpr::getAdd(RC, Scale);
5208 // Emit an add instruction.
5209 Result = IC.InsertNewInstBefore(
5210 BinaryOperator::CreateAdd(Result, Scale,
5211 GEP->getName()+".offs"), I);
5215 // Convert to correct type.
5216 if (Op->getType() != IntPtrTy) {
5217 if (Constant *OpC = dyn_cast<Constant>(Op))
5218 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
5220 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
5221 Op->getName()+".c"), I);
5224 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5225 if (Constant *OpC = dyn_cast<Constant>(Op))
5226 Op = ConstantExpr::getMul(OpC, Scale);
5227 else // We'll let instcombine(mul) convert this to a shl if possible.
5228 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
5229 GEP->getName()+".idx"), I);
5232 // Emit an add instruction.
5233 if (isa<Constant>(Op) && isa<Constant>(Result))
5234 Result = ConstantExpr::getAdd(cast<Constant>(Op),
5235 cast<Constant>(Result));
5237 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
5238 GEP->getName()+".offs"), I);
5244 /// EvaluateGEPOffsetExpression - Return an value that can be used to compare of
5245 /// the *offset* implied by GEP to zero. For example, if we have &A[i], we want
5246 /// to return 'i' for "icmp ne i, 0". Note that, in general, indices can be
5247 /// complex, and scales are involved. The above expression would also be legal
5248 /// to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). This
5249 /// later form is less amenable to optimization though, and we are allowed to
5250 /// generate the first by knowing that pointer arithmetic doesn't overflow.
5252 /// If we can't emit an optimized form for this expression, this returns null.
5254 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
5256 TargetData &TD = IC.getTargetData();
5257 gep_type_iterator GTI = gep_type_begin(GEP);
5259 // Check to see if this gep only has a single variable index. If so, and if
5260 // any constant indices are a multiple of its scale, then we can compute this
5261 // in terms of the scale of the variable index. For example, if the GEP
5262 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
5263 // because the expression will cross zero at the same point.
5264 unsigned i, e = GEP->getNumOperands();
5266 for (i = 1; i != e; ++i, ++GTI) {
5267 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
5268 // Compute the aggregate offset of constant indices.
5269 if (CI->isZero()) continue;
5271 // Handle a struct index, which adds its field offset to the pointer.
5272 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5273 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5275 uint64_t Size = TD.getTypePaddedSize(GTI.getIndexedType());
5276 Offset += Size*CI->getSExtValue();
5279 // Found our variable index.
5284 // If there are no variable indices, we must have a constant offset, just
5285 // evaluate it the general way.
5286 if (i == e) return 0;
5288 Value *VariableIdx = GEP->getOperand(i);
5289 // Determine the scale factor of the variable element. For example, this is
5290 // 4 if the variable index is into an array of i32.
5291 uint64_t VariableScale = TD.getTypePaddedSize(GTI.getIndexedType());
5293 // Verify that there are no other variable indices. If so, emit the hard way.
5294 for (++i, ++GTI; i != e; ++i, ++GTI) {
5295 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
5298 // Compute the aggregate offset of constant indices.
5299 if (CI->isZero()) continue;
5301 // Handle a struct index, which adds its field offset to the pointer.
5302 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5303 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5305 uint64_t Size = TD.getTypePaddedSize(GTI.getIndexedType());
5306 Offset += Size*CI->getSExtValue();
5310 // Okay, we know we have a single variable index, which must be a
5311 // pointer/array/vector index. If there is no offset, life is simple, return
5313 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5315 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5316 // we don't need to bother extending: the extension won't affect where the
5317 // computation crosses zero.
5318 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5319 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
5320 VariableIdx->getNameStart(), &I);
5324 // Otherwise, there is an index. The computation we will do will be modulo
5325 // the pointer size, so get it.
5326 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5328 Offset &= PtrSizeMask;
5329 VariableScale &= PtrSizeMask;
5331 // To do this transformation, any constant index must be a multiple of the
5332 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5333 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5334 // multiple of the variable scale.
5335 int64_t NewOffs = Offset / (int64_t)VariableScale;
5336 if (Offset != NewOffs*(int64_t)VariableScale)
5339 // Okay, we can do this evaluation. Start by converting the index to intptr.
5340 const Type *IntPtrTy = TD.getIntPtrType();
5341 if (VariableIdx->getType() != IntPtrTy)
5342 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
5344 VariableIdx->getNameStart(), &I);
5345 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
5346 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
5350 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5351 /// else. At this point we know that the GEP is on the LHS of the comparison.
5352 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
5353 ICmpInst::Predicate Cond,
5355 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
5357 // Look through bitcasts.
5358 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5359 RHS = BCI->getOperand(0);
5361 Value *PtrBase = GEPLHS->getOperand(0);
5362 if (PtrBase == RHS) {
5363 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5364 // This transformation (ignoring the base and scales) is valid because we
5365 // know pointers can't overflow. See if we can output an optimized form.
5366 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5368 // If not, synthesize the offset the hard way.
5370 Offset = EmitGEPOffset(GEPLHS, I, *this);
5371 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
5372 Constant::getNullValue(Offset->getType()));
5373 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
5374 // If the base pointers are different, but the indices are the same, just
5375 // compare the base pointer.
5376 if (PtrBase != GEPRHS->getOperand(0)) {
5377 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5378 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5379 GEPRHS->getOperand(0)->getType();
5381 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5382 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5383 IndicesTheSame = false;
5387 // If all indices are the same, just compare the base pointers.
5389 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
5390 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5392 // Otherwise, the base pointers are different and the indices are
5393 // different, bail out.
5397 // If one of the GEPs has all zero indices, recurse.
5398 bool AllZeros = true;
5399 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5400 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5401 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5406 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5407 ICmpInst::getSwappedPredicate(Cond), I);
5409 // If the other GEP has all zero indices, recurse.
5411 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5412 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5413 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5418 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5420 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5421 // If the GEPs only differ by one index, compare it.
5422 unsigned NumDifferences = 0; // Keep track of # differences.
5423 unsigned DiffOperand = 0; // The operand that differs.
5424 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5425 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5426 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5427 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5428 // Irreconcilable differences.
5432 if (NumDifferences++) break;
5437 if (NumDifferences == 0) // SAME GEP?
5438 return ReplaceInstUsesWith(I, // No comparison is needed here.
5439 ConstantInt::get(Type::Int1Ty,
5440 ICmpInst::isTrueWhenEqual(Cond)));
5442 else if (NumDifferences == 1) {
5443 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5444 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5445 // Make sure we do a signed comparison here.
5446 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5450 // Only lower this if the icmp is the only user of the GEP or if we expect
5451 // the result to fold to a constant!
5452 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5453 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5454 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5455 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5456 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5457 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
5463 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5465 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5468 if (!isa<ConstantFP>(RHSC)) return 0;
5469 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5471 // Get the width of the mantissa. We don't want to hack on conversions that
5472 // might lose information from the integer, e.g. "i64 -> float"
5473 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5474 if (MantissaWidth == -1) return 0; // Unknown.
5476 // Check to see that the input is converted from an integer type that is small
5477 // enough that preserves all bits. TODO: check here for "known" sign bits.
5478 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5479 unsigned InputSize = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
5481 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5482 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5486 // If the conversion would lose info, don't hack on this.
5487 if ((int)InputSize > MantissaWidth)
5490 // Otherwise, we can potentially simplify the comparison. We know that it
5491 // will always come through as an integer value and we know the constant is
5492 // not a NAN (it would have been previously simplified).
5493 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5495 ICmpInst::Predicate Pred;
5496 switch (I.getPredicate()) {
5497 default: assert(0 && "Unexpected predicate!");
5498 case FCmpInst::FCMP_UEQ:
5499 case FCmpInst::FCMP_OEQ:
5500 Pred = ICmpInst::ICMP_EQ;
5502 case FCmpInst::FCMP_UGT:
5503 case FCmpInst::FCMP_OGT:
5504 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5506 case FCmpInst::FCMP_UGE:
5507 case FCmpInst::FCMP_OGE:
5508 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5510 case FCmpInst::FCMP_ULT:
5511 case FCmpInst::FCMP_OLT:
5512 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5514 case FCmpInst::FCMP_ULE:
5515 case FCmpInst::FCMP_OLE:
5516 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5518 case FCmpInst::FCMP_UNE:
5519 case FCmpInst::FCMP_ONE:
5520 Pred = ICmpInst::ICMP_NE;
5522 case FCmpInst::FCMP_ORD:
5523 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5524 case FCmpInst::FCMP_UNO:
5525 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5528 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5530 // Now we know that the APFloat is a normal number, zero or inf.
5532 // See if the FP constant is too large for the integer. For example,
5533 // comparing an i8 to 300.0.
5534 unsigned IntWidth = IntTy->getPrimitiveSizeInBits();
5537 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5538 // and large values.
5539 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5540 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5541 APFloat::rmNearestTiesToEven);
5542 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5543 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5544 Pred == ICmpInst::ICMP_SLE)
5545 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5546 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5549 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5550 // +INF and large values.
5551 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
5552 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5553 APFloat::rmNearestTiesToEven);
5554 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5555 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5556 Pred == ICmpInst::ICMP_ULE)
5557 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5558 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5563 // See if the RHS value is < SignedMin.
5564 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5565 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5566 APFloat::rmNearestTiesToEven);
5567 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5568 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5569 Pred == ICmpInst::ICMP_SGE)
5570 return ReplaceInstUsesWith(I,ConstantInt::getTrue());
5571 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5575 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5576 // [0, UMAX], but it may still be fractional. See if it is fractional by
5577 // casting the FP value to the integer value and back, checking for equality.
5578 // Don't do this for zero, because -0.0 is not fractional.
5579 Constant *RHSInt = ConstantExpr::getFPToSI(RHSC, IntTy);
5580 if (!RHS.isZero() &&
5581 ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) != RHSC) {
5582 // If we had a comparison against a fractional value, we have to adjust the
5583 // compare predicate and sometimes the value. RHSC is rounded towards zero
5586 default: assert(0 && "Unexpected integer comparison!");
5587 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5588 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5589 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5590 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5591 case ICmpInst::ICMP_ULE:
5592 // (float)int <= 4.4 --> int <= 4
5593 // (float)int <= -4.4 --> false
5594 if (RHS.isNegative())
5595 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5597 case ICmpInst::ICMP_SLE:
5598 // (float)int <= 4.4 --> int <= 4
5599 // (float)int <= -4.4 --> int < -4
5600 if (RHS.isNegative())
5601 Pred = ICmpInst::ICMP_SLT;
5603 case ICmpInst::ICMP_ULT:
5604 // (float)int < -4.4 --> false
5605 // (float)int < 4.4 --> int <= 4
5606 if (RHS.isNegative())
5607 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5608 Pred = ICmpInst::ICMP_ULE;
5610 case ICmpInst::ICMP_SLT:
5611 // (float)int < -4.4 --> int < -4
5612 // (float)int < 4.4 --> int <= 4
5613 if (!RHS.isNegative())
5614 Pred = ICmpInst::ICMP_SLE;
5616 case ICmpInst::ICMP_UGT:
5617 // (float)int > 4.4 --> int > 4
5618 // (float)int > -4.4 --> true
5619 if (RHS.isNegative())
5620 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5622 case ICmpInst::ICMP_SGT:
5623 // (float)int > 4.4 --> int > 4
5624 // (float)int > -4.4 --> int >= -4
5625 if (RHS.isNegative())
5626 Pred = ICmpInst::ICMP_SGE;
5628 case ICmpInst::ICMP_UGE:
5629 // (float)int >= -4.4 --> true
5630 // (float)int >= 4.4 --> int > 4
5631 if (!RHS.isNegative())
5632 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5633 Pred = ICmpInst::ICMP_UGT;
5635 case ICmpInst::ICMP_SGE:
5636 // (float)int >= -4.4 --> int >= -4
5637 // (float)int >= 4.4 --> int > 4
5638 if (!RHS.isNegative())
5639 Pred = ICmpInst::ICMP_SGT;
5644 // Lower this FP comparison into an appropriate integer version of the
5646 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5649 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5650 bool Changed = SimplifyCompare(I);
5651 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5653 // Fold trivial predicates.
5654 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5655 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5656 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5657 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5659 // Simplify 'fcmp pred X, X'
5661 switch (I.getPredicate()) {
5662 default: assert(0 && "Unknown predicate!");
5663 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5664 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5665 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5666 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5667 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5668 case FCmpInst::FCMP_OLT: // True if ordered and less than
5669 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5670 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5672 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5673 case FCmpInst::FCMP_ULT: // True if unordered or less than
5674 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5675 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5676 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5677 I.setPredicate(FCmpInst::FCMP_UNO);
5678 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5681 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5682 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5683 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5684 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5685 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5686 I.setPredicate(FCmpInst::FCMP_ORD);
5687 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5692 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5693 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5695 // Handle fcmp with constant RHS
5696 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5697 // If the constant is a nan, see if we can fold the comparison based on it.
5698 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5699 if (CFP->getValueAPF().isNaN()) {
5700 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5701 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5702 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5703 "Comparison must be either ordered or unordered!");
5704 // True if unordered.
5705 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5709 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5710 switch (LHSI->getOpcode()) {
5711 case Instruction::PHI:
5712 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5713 // block. If in the same block, we're encouraging jump threading. If
5714 // not, we are just pessimizing the code by making an i1 phi.
5715 if (LHSI->getParent() == I.getParent())
5716 if (Instruction *NV = FoldOpIntoPhi(I))
5719 case Instruction::SIToFP:
5720 case Instruction::UIToFP:
5721 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5724 case Instruction::Select:
5725 // If either operand of the select is a constant, we can fold the
5726 // comparison into the select arms, which will cause one to be
5727 // constant folded and the select turned into a bitwise or.
5728 Value *Op1 = 0, *Op2 = 0;
5729 if (LHSI->hasOneUse()) {
5730 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5731 // Fold the known value into the constant operand.
5732 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5733 // Insert a new FCmp of the other select operand.
5734 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5735 LHSI->getOperand(2), RHSC,
5737 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5738 // Fold the known value into the constant operand.
5739 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5740 // Insert a new FCmp of the other select operand.
5741 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5742 LHSI->getOperand(1), RHSC,
5748 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5753 return Changed ? &I : 0;
5756 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5757 bool Changed = SimplifyCompare(I);
5758 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5759 const Type *Ty = Op0->getType();
5763 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5764 I.isTrueWhenEqual()));
5766 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5767 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5769 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5770 // addresses never equal each other! We already know that Op0 != Op1.
5771 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5772 isa<ConstantPointerNull>(Op0)) &&
5773 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5774 isa<ConstantPointerNull>(Op1)))
5775 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5776 !I.isTrueWhenEqual()));
5778 // icmp's with boolean values can always be turned into bitwise operations
5779 if (Ty == Type::Int1Ty) {
5780 switch (I.getPredicate()) {
5781 default: assert(0 && "Invalid icmp instruction!");
5782 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
5783 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
5784 InsertNewInstBefore(Xor, I);
5785 return BinaryOperator::CreateNot(Xor);
5787 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
5788 return BinaryOperator::CreateXor(Op0, Op1);
5790 case ICmpInst::ICMP_UGT:
5791 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
5793 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
5794 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5795 InsertNewInstBefore(Not, I);
5796 return BinaryOperator::CreateAnd(Not, Op1);
5798 case ICmpInst::ICMP_SGT:
5799 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
5801 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
5802 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5803 InsertNewInstBefore(Not, I);
5804 return BinaryOperator::CreateAnd(Not, Op0);
5806 case ICmpInst::ICMP_UGE:
5807 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
5809 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
5810 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
5811 InsertNewInstBefore(Not, I);
5812 return BinaryOperator::CreateOr(Not, Op1);
5814 case ICmpInst::ICMP_SGE:
5815 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
5817 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
5818 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
5819 InsertNewInstBefore(Not, I);
5820 return BinaryOperator::CreateOr(Not, Op0);
5825 // See if we are doing a comparison with a constant.
5826 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5827 Value *A = 0, *B = 0;
5829 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
5830 if (I.isEquality() && CI->isNullValue() &&
5831 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
5832 // (icmp cond A B) if cond is equality
5833 return new ICmpInst(I.getPredicate(), A, B);
5836 // If we have an icmp le or icmp ge instruction, turn it into the
5837 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
5838 // them being folded in the code below.
5839 switch (I.getPredicate()) {
5841 case ICmpInst::ICMP_ULE:
5842 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
5843 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5844 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5845 case ICmpInst::ICMP_SLE:
5846 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
5847 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5848 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5849 case ICmpInst::ICMP_UGE:
5850 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
5851 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5852 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5853 case ICmpInst::ICMP_SGE:
5854 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5855 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5856 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5859 // See if we can fold the comparison based on range information we can get
5860 // by checking whether bits are known to be zero or one in the input.
5861 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5862 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5864 // If this comparison is a normal comparison, it demands all
5865 // bits, if it is a sign bit comparison, it only demands the sign bit.
5867 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5869 if (SimplifyDemandedBits(I.getOperandUse(0),
5870 isSignBit ? APInt::getSignBit(BitWidth)
5871 : APInt::getAllOnesValue(BitWidth),
5872 KnownZero, KnownOne, 0))
5875 // Given the known and unknown bits, compute a range that the LHS could be
5876 // in. Compute the Min, Max and RHS values based on the known bits. For the
5877 // EQ and NE we use unsigned values.
5878 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5879 if (ICmpInst::isSignedPredicate(I.getPredicate()))
5880 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min, Max);
5882 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,Min,Max);
5884 // If Min and Max are known to be the same, then SimplifyDemandedBits
5885 // figured out that the LHS is a constant. Just constant fold this now so
5886 // that code below can assume that Min != Max.
5888 return ReplaceInstUsesWith(I, ConstantExpr::getICmp(I.getPredicate(),
5889 ConstantInt::get(Min),
5892 // Based on the range information we know about the LHS, see if we can
5893 // simplify this comparison. For example, (x&4) < 8 is always true.
5894 const APInt &RHSVal = CI->getValue();
5895 switch (I.getPredicate()) { // LE/GE have been folded already.
5896 default: assert(0 && "Unknown icmp opcode!");
5897 case ICmpInst::ICMP_EQ:
5898 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5899 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5901 case ICmpInst::ICMP_NE:
5902 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5903 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5905 case ICmpInst::ICMP_ULT:
5906 if (Max.ult(RHSVal)) // A <u C -> true iff max(A) < C
5907 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5908 if (Min.uge(RHSVal)) // A <u C -> false iff min(A) >= C
5909 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5910 if (RHSVal == Max) // A <u MAX -> A != MAX
5911 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5912 if (RHSVal == Min+1) // A <u MIN+1 -> A == MIN
5913 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5915 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
5916 if (CI->isMinValue(true))
5917 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
5918 ConstantInt::getAllOnesValue(Op0->getType()));
5920 case ICmpInst::ICMP_UGT:
5921 if (Min.ugt(RHSVal)) // A >u C -> true iff min(A) > C
5922 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5923 if (Max.ule(RHSVal)) // A >u C -> false iff max(A) <= C
5924 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5926 if (RHSVal == Min) // A >u MIN -> A != MIN
5927 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5928 if (RHSVal == Max-1) // A >u MAX-1 -> A == MAX
5929 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5931 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
5932 if (CI->isMaxValue(true))
5933 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
5934 ConstantInt::getNullValue(Op0->getType()));
5936 case ICmpInst::ICMP_SLT:
5937 if (Max.slt(RHSVal)) // A <s C -> true iff max(A) < C
5938 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5939 if (Min.sge(RHSVal)) // A <s C -> false iff min(A) >= C
5940 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5941 if (RHSVal == Max) // A <s MAX -> A != MAX
5942 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5943 if (RHSVal == Min+1) // A <s MIN+1 -> A == MIN
5944 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5946 case ICmpInst::ICMP_SGT:
5947 if (Min.sgt(RHSVal)) // A >s C -> true iff min(A) > C
5948 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5949 if (Max.sle(RHSVal)) // A >s C -> false iff max(A) <= C
5950 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5952 if (RHSVal == Min) // A >s MIN -> A != MIN
5953 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5954 if (RHSVal == Max-1) // A >s MAX-1 -> A == MAX
5955 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5960 // Test if the ICmpInst instruction is used exclusively by a select as
5961 // part of a minimum or maximum operation. If so, refrain from doing
5962 // any other folding. This helps out other analyses which understand
5963 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
5964 // and CodeGen. And in this case, at least one of the comparison
5965 // operands has at least one user besides the compare (the select),
5966 // which would often largely negate the benefit of folding anyway.
5968 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
5969 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
5970 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
5973 // See if we are doing a comparison between a constant and an instruction that
5974 // can be folded into the comparison.
5975 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5976 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5977 // instruction, see if that instruction also has constants so that the
5978 // instruction can be folded into the icmp
5979 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5980 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5984 // Handle icmp with constant (but not simple integer constant) RHS
5985 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5986 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5987 switch (LHSI->getOpcode()) {
5988 case Instruction::GetElementPtr:
5989 if (RHSC->isNullValue()) {
5990 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5991 bool isAllZeros = true;
5992 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5993 if (!isa<Constant>(LHSI->getOperand(i)) ||
5994 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5999 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
6000 Constant::getNullValue(LHSI->getOperand(0)->getType()));
6004 case Instruction::PHI:
6005 // Only fold icmp into the PHI if the phi and fcmp are in the same
6006 // block. If in the same block, we're encouraging jump threading. If
6007 // not, we are just pessimizing the code by making an i1 phi.
6008 if (LHSI->getParent() == I.getParent())
6009 if (Instruction *NV = FoldOpIntoPhi(I))
6012 case Instruction::Select: {
6013 // If either operand of the select is a constant, we can fold the
6014 // comparison into the select arms, which will cause one to be
6015 // constant folded and the select turned into a bitwise or.
6016 Value *Op1 = 0, *Op2 = 0;
6017 if (LHSI->hasOneUse()) {
6018 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
6019 // Fold the known value into the constant operand.
6020 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6021 // Insert a new ICmp of the other select operand.
6022 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
6023 LHSI->getOperand(2), RHSC,
6025 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
6026 // Fold the known value into the constant operand.
6027 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6028 // Insert a new ICmp of the other select operand.
6029 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
6030 LHSI->getOperand(1), RHSC,
6036 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
6039 case Instruction::Malloc:
6040 // If we have (malloc != null), and if the malloc has a single use, we
6041 // can assume it is successful and remove the malloc.
6042 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
6043 AddToWorkList(LHSI);
6044 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
6045 !I.isTrueWhenEqual()));
6051 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
6052 if (User *GEP = dyn_castGetElementPtr(Op0))
6053 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
6055 if (User *GEP = dyn_castGetElementPtr(Op1))
6056 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
6057 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
6060 // Test to see if the operands of the icmp are casted versions of other
6061 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
6063 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
6064 if (isa<PointerType>(Op0->getType()) &&
6065 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
6066 // We keep moving the cast from the left operand over to the right
6067 // operand, where it can often be eliminated completely.
6068 Op0 = CI->getOperand(0);
6070 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
6071 // so eliminate it as well.
6072 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
6073 Op1 = CI2->getOperand(0);
6075 // If Op1 is a constant, we can fold the cast into the constant.
6076 if (Op0->getType() != Op1->getType()) {
6077 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
6078 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
6080 // Otherwise, cast the RHS right before the icmp
6081 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
6084 return new ICmpInst(I.getPredicate(), Op0, Op1);
6088 if (isa<CastInst>(Op0)) {
6089 // Handle the special case of: icmp (cast bool to X), <cst>
6090 // This comes up when you have code like
6093 // For generality, we handle any zero-extension of any operand comparison
6094 // with a constant or another cast from the same type.
6095 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
6096 if (Instruction *R = visitICmpInstWithCastAndCast(I))
6100 // See if it's the same type of instruction on the left and right.
6101 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
6102 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
6103 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
6104 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1)) {
6105 switch (Op0I->getOpcode()) {
6107 case Instruction::Add:
6108 case Instruction::Sub:
6109 case Instruction::Xor:
6110 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
6111 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
6112 Op1I->getOperand(0));
6113 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
6114 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6115 if (CI->getValue().isSignBit()) {
6116 ICmpInst::Predicate Pred = I.isSignedPredicate()
6117 ? I.getUnsignedPredicate()
6118 : I.getSignedPredicate();
6119 return new ICmpInst(Pred, Op0I->getOperand(0),
6120 Op1I->getOperand(0));
6123 if (CI->getValue().isMaxSignedValue()) {
6124 ICmpInst::Predicate Pred = I.isSignedPredicate()
6125 ? I.getUnsignedPredicate()
6126 : I.getSignedPredicate();
6127 Pred = I.getSwappedPredicate(Pred);
6128 return new ICmpInst(Pred, Op0I->getOperand(0),
6129 Op1I->getOperand(0));
6133 case Instruction::Mul:
6134 if (!I.isEquality())
6137 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6138 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
6139 // Mask = -1 >> count-trailing-zeros(Cst).
6140 if (!CI->isZero() && !CI->isOne()) {
6141 const APInt &AP = CI->getValue();
6142 ConstantInt *Mask = ConstantInt::get(
6143 APInt::getLowBitsSet(AP.getBitWidth(),
6145 AP.countTrailingZeros()));
6146 Instruction *And1 = BinaryOperator::CreateAnd(Op0I->getOperand(0),
6148 Instruction *And2 = BinaryOperator::CreateAnd(Op1I->getOperand(0),
6150 InsertNewInstBefore(And1, I);
6151 InsertNewInstBefore(And2, I);
6152 return new ICmpInst(I.getPredicate(), And1, And2);
6161 // ~x < ~y --> y < x
6163 if (match(Op0, m_Not(m_Value(A))) &&
6164 match(Op1, m_Not(m_Value(B))))
6165 return new ICmpInst(I.getPredicate(), B, A);
6168 if (I.isEquality()) {
6169 Value *A, *B, *C, *D;
6171 // -x == -y --> x == y
6172 if (match(Op0, m_Neg(m_Value(A))) &&
6173 match(Op1, m_Neg(m_Value(B))))
6174 return new ICmpInst(I.getPredicate(), A, B);
6176 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
6177 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6178 Value *OtherVal = A == Op1 ? B : A;
6179 return new ICmpInst(I.getPredicate(), OtherVal,
6180 Constant::getNullValue(A->getType()));
6183 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
6184 // A^c1 == C^c2 --> A == C^(c1^c2)
6185 ConstantInt *C1, *C2;
6186 if (match(B, m_ConstantInt(C1)) &&
6187 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
6188 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
6189 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
6190 return new ICmpInst(I.getPredicate(), A,
6191 InsertNewInstBefore(Xor, I));
6194 // A^B == A^D -> B == D
6195 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
6196 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
6197 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
6198 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
6202 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
6203 (A == Op0 || B == Op0)) {
6204 // A == (A^B) -> B == 0
6205 Value *OtherVal = A == Op0 ? B : A;
6206 return new ICmpInst(I.getPredicate(), OtherVal,
6207 Constant::getNullValue(A->getType()));
6210 // (A-B) == A -> B == 0
6211 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B))))
6212 return new ICmpInst(I.getPredicate(), B,
6213 Constant::getNullValue(B->getType()));
6215 // A == (A-B) -> B == 0
6216 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B))))
6217 return new ICmpInst(I.getPredicate(), B,
6218 Constant::getNullValue(B->getType()));
6220 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6221 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6222 match(Op0, m_And(m_Value(A), m_Value(B))) &&
6223 match(Op1, m_And(m_Value(C), m_Value(D)))) {
6224 Value *X = 0, *Y = 0, *Z = 0;
6227 X = B; Y = D; Z = A;
6228 } else if (A == D) {
6229 X = B; Y = C; Z = A;
6230 } else if (B == C) {
6231 X = A; Y = D; Z = B;
6232 } else if (B == D) {
6233 X = A; Y = C; Z = B;
6236 if (X) { // Build (X^Y) & Z
6237 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
6238 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
6239 I.setOperand(0, Op1);
6240 I.setOperand(1, Constant::getNullValue(Op1->getType()));
6245 return Changed ? &I : 0;
6249 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6250 /// and CmpRHS are both known to be integer constants.
6251 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6252 ConstantInt *DivRHS) {
6253 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6254 const APInt &CmpRHSV = CmpRHS->getValue();
6256 // FIXME: If the operand types don't match the type of the divide
6257 // then don't attempt this transform. The code below doesn't have the
6258 // logic to deal with a signed divide and an unsigned compare (and
6259 // vice versa). This is because (x /s C1) <s C2 produces different
6260 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6261 // (x /u C1) <u C2. Simply casting the operands and result won't
6262 // work. :( The if statement below tests that condition and bails
6264 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6265 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
6267 if (DivRHS->isZero())
6268 return 0; // The ProdOV computation fails on divide by zero.
6269 if (DivIsSigned && DivRHS->isAllOnesValue())
6270 return 0; // The overflow computation also screws up here
6271 if (DivRHS->isOne())
6272 return 0; // Not worth bothering, and eliminates some funny cases
6275 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6276 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6277 // C2 (CI). By solving for X we can turn this into a range check
6278 // instead of computing a divide.
6279 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
6281 // Determine if the product overflows by seeing if the product is
6282 // not equal to the divide. Make sure we do the same kind of divide
6283 // as in the LHS instruction that we're folding.
6284 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
6285 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
6287 // Get the ICmp opcode
6288 ICmpInst::Predicate Pred = ICI.getPredicate();
6290 // Figure out the interval that is being checked. For example, a comparison
6291 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
6292 // Compute this interval based on the constants involved and the signedness of
6293 // the compare/divide. This computes a half-open interval, keeping track of
6294 // whether either value in the interval overflows. After analysis each
6295 // overflow variable is set to 0 if it's corresponding bound variable is valid
6296 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
6297 int LoOverflow = 0, HiOverflow = 0;
6298 ConstantInt *LoBound = 0, *HiBound = 0;
6300 if (!DivIsSigned) { // udiv
6301 // e.g. X/5 op 3 --> [15, 20)
6303 HiOverflow = LoOverflow = ProdOV;
6305 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
6306 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
6307 if (CmpRHSV == 0) { // (X / pos) op 0
6308 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
6309 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
6311 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
6312 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
6313 HiOverflow = LoOverflow = ProdOV;
6315 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
6316 } else { // (X / pos) op neg
6317 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
6318 HiBound = AddOne(Prod);
6319 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
6321 ConstantInt* DivNeg = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6322 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg,
6326 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
6327 if (CmpRHSV == 0) { // (X / neg) op 0
6328 // e.g. X/-5 op 0 --> [-4, 5)
6329 LoBound = AddOne(DivRHS);
6330 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6331 if (HiBound == DivRHS) { // -INTMIN = INTMIN
6332 HiOverflow = 1; // [INTMIN+1, overflow)
6333 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
6335 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
6336 // e.g. X/-5 op 3 --> [-19, -14)
6337 HiBound = AddOne(Prod);
6338 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
6340 LoOverflow = AddWithOverflow(LoBound, HiBound, DivRHS, true) ? -1 : 0;
6341 } else { // (X / neg) op neg
6342 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
6343 LoOverflow = HiOverflow = ProdOV;
6345 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, true);
6348 // Dividing by a negative swaps the condition. LT <-> GT
6349 Pred = ICmpInst::getSwappedPredicate(Pred);
6352 Value *X = DivI->getOperand(0);
6354 default: assert(0 && "Unhandled icmp opcode!");
6355 case ICmpInst::ICMP_EQ:
6356 if (LoOverflow && HiOverflow)
6357 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6358 else if (HiOverflow)
6359 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6360 ICmpInst::ICMP_UGE, X, LoBound);
6361 else if (LoOverflow)
6362 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6363 ICmpInst::ICMP_ULT, X, HiBound);
6365 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
6366 case ICmpInst::ICMP_NE:
6367 if (LoOverflow && HiOverflow)
6368 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6369 else if (HiOverflow)
6370 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6371 ICmpInst::ICMP_ULT, X, LoBound);
6372 else if (LoOverflow)
6373 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6374 ICmpInst::ICMP_UGE, X, HiBound);
6376 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
6377 case ICmpInst::ICMP_ULT:
6378 case ICmpInst::ICMP_SLT:
6379 if (LoOverflow == +1) // Low bound is greater than input range.
6380 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6381 if (LoOverflow == -1) // Low bound is less than input range.
6382 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6383 return new ICmpInst(Pred, X, LoBound);
6384 case ICmpInst::ICMP_UGT:
6385 case ICmpInst::ICMP_SGT:
6386 if (HiOverflow == +1) // High bound greater than input range.
6387 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6388 else if (HiOverflow == -1) // High bound less than input range.
6389 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6390 if (Pred == ICmpInst::ICMP_UGT)
6391 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
6393 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
6398 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
6400 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
6403 const APInt &RHSV = RHS->getValue();
6405 switch (LHSI->getOpcode()) {
6406 case Instruction::Trunc:
6407 if (ICI.isEquality() && LHSI->hasOneUse()) {
6408 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
6409 // of the high bits truncated out of x are known.
6410 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
6411 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
6412 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
6413 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
6414 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
6416 // If all the high bits are known, we can do this xform.
6417 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
6418 // Pull in the high bits from known-ones set.
6419 APInt NewRHS(RHS->getValue());
6420 NewRHS.zext(SrcBits);
6422 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6423 ConstantInt::get(NewRHS));
6428 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
6429 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
6430 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
6432 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
6433 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
6434 Value *CompareVal = LHSI->getOperand(0);
6436 // If the sign bit of the XorCST is not set, there is no change to
6437 // the operation, just stop using the Xor.
6438 if (!XorCST->getValue().isNegative()) {
6439 ICI.setOperand(0, CompareVal);
6440 AddToWorkList(LHSI);
6444 // Was the old condition true if the operand is positive?
6445 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
6447 // If so, the new one isn't.
6448 isTrueIfPositive ^= true;
6450 if (isTrueIfPositive)
6451 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
6453 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
6456 if (LHSI->hasOneUse()) {
6457 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
6458 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
6459 const APInt &SignBit = XorCST->getValue();
6460 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6461 ? ICI.getUnsignedPredicate()
6462 : ICI.getSignedPredicate();
6463 return new ICmpInst(Pred, LHSI->getOperand(0),
6464 ConstantInt::get(RHSV ^ SignBit));
6467 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
6468 if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) {
6469 const APInt &NotSignBit = XorCST->getValue();
6470 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6471 ? ICI.getUnsignedPredicate()
6472 : ICI.getSignedPredicate();
6473 Pred = ICI.getSwappedPredicate(Pred);
6474 return new ICmpInst(Pred, LHSI->getOperand(0),
6475 ConstantInt::get(RHSV ^ NotSignBit));
6480 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
6481 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
6482 LHSI->getOperand(0)->hasOneUse()) {
6483 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
6485 // If the LHS is an AND of a truncating cast, we can widen the
6486 // and/compare to be the input width without changing the value
6487 // produced, eliminating a cast.
6488 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
6489 // We can do this transformation if either the AND constant does not
6490 // have its sign bit set or if it is an equality comparison.
6491 // Extending a relational comparison when we're checking the sign
6492 // bit would not work.
6493 if (Cast->hasOneUse() &&
6494 (ICI.isEquality() ||
6495 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
6497 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
6498 APInt NewCST = AndCST->getValue();
6499 NewCST.zext(BitWidth);
6501 NewCI.zext(BitWidth);
6502 Instruction *NewAnd =
6503 BinaryOperator::CreateAnd(Cast->getOperand(0),
6504 ConstantInt::get(NewCST),LHSI->getName());
6505 InsertNewInstBefore(NewAnd, ICI);
6506 return new ICmpInst(ICI.getPredicate(), NewAnd,
6507 ConstantInt::get(NewCI));
6511 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
6512 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
6513 // happens a LOT in code produced by the C front-end, for bitfield
6515 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
6516 if (Shift && !Shift->isShift())
6520 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
6521 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
6522 const Type *AndTy = AndCST->getType(); // Type of the and.
6524 // We can fold this as long as we can't shift unknown bits
6525 // into the mask. This can only happen with signed shift
6526 // rights, as they sign-extend.
6528 bool CanFold = Shift->isLogicalShift();
6530 // To test for the bad case of the signed shr, see if any
6531 // of the bits shifted in could be tested after the mask.
6532 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
6533 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
6535 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
6536 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
6537 AndCST->getValue()) == 0)
6543 if (Shift->getOpcode() == Instruction::Shl)
6544 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
6546 NewCst = ConstantExpr::getShl(RHS, ShAmt);
6548 // Check to see if we are shifting out any of the bits being
6550 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
6551 // If we shifted bits out, the fold is not going to work out.
6552 // As a special case, check to see if this means that the
6553 // result is always true or false now.
6554 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6555 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6556 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6557 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6559 ICI.setOperand(1, NewCst);
6560 Constant *NewAndCST;
6561 if (Shift->getOpcode() == Instruction::Shl)
6562 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
6564 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
6565 LHSI->setOperand(1, NewAndCST);
6566 LHSI->setOperand(0, Shift->getOperand(0));
6567 AddToWorkList(Shift); // Shift is dead.
6568 AddUsesToWorkList(ICI);
6574 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
6575 // preferable because it allows the C<<Y expression to be hoisted out
6576 // of a loop if Y is invariant and X is not.
6577 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
6578 ICI.isEquality() && !Shift->isArithmeticShift() &&
6579 isa<Instruction>(Shift->getOperand(0))) {
6582 if (Shift->getOpcode() == Instruction::LShr) {
6583 NS = BinaryOperator::CreateShl(AndCST,
6584 Shift->getOperand(1), "tmp");
6586 // Insert a logical shift.
6587 NS = BinaryOperator::CreateLShr(AndCST,
6588 Shift->getOperand(1), "tmp");
6590 InsertNewInstBefore(cast<Instruction>(NS), ICI);
6592 // Compute X & (C << Y).
6593 Instruction *NewAnd =
6594 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
6595 InsertNewInstBefore(NewAnd, ICI);
6597 ICI.setOperand(0, NewAnd);
6603 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6604 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6607 uint32_t TypeBits = RHSV.getBitWidth();
6609 // Check that the shift amount is in range. If not, don't perform
6610 // undefined shifts. When the shift is visited it will be
6612 if (ShAmt->uge(TypeBits))
6615 if (ICI.isEquality()) {
6616 // If we are comparing against bits always shifted out, the
6617 // comparison cannot succeed.
6619 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
6620 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6621 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6622 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6623 return ReplaceInstUsesWith(ICI, Cst);
6626 if (LHSI->hasOneUse()) {
6627 // Otherwise strength reduce the shift into an and.
6628 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6630 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
6633 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6634 Mask, LHSI->getName()+".mask");
6635 Value *And = InsertNewInstBefore(AndI, ICI);
6636 return new ICmpInst(ICI.getPredicate(), And,
6637 ConstantInt::get(RHSV.lshr(ShAmtVal)));
6641 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6642 bool TrueIfSigned = false;
6643 if (LHSI->hasOneUse() &&
6644 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6645 // (X << 31) <s 0 --> (X&1) != 0
6646 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
6647 (TypeBits-ShAmt->getZExtValue()-1));
6649 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6650 Mask, LHSI->getName()+".mask");
6651 Value *And = InsertNewInstBefore(AndI, ICI);
6653 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6654 And, Constant::getNullValue(And->getType()));
6659 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6660 case Instruction::AShr: {
6661 // Only handle equality comparisons of shift-by-constant.
6662 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6663 if (!ShAmt || !ICI.isEquality()) break;
6665 // Check that the shift amount is in range. If not, don't perform
6666 // undefined shifts. When the shift is visited it will be
6668 uint32_t TypeBits = RHSV.getBitWidth();
6669 if (ShAmt->uge(TypeBits))
6672 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6674 // If we are comparing against bits always shifted out, the
6675 // comparison cannot succeed.
6676 APInt Comp = RHSV << ShAmtVal;
6677 if (LHSI->getOpcode() == Instruction::LShr)
6678 Comp = Comp.lshr(ShAmtVal);
6680 Comp = Comp.ashr(ShAmtVal);
6682 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6683 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6684 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6685 return ReplaceInstUsesWith(ICI, Cst);
6688 // Otherwise, check to see if the bits shifted out are known to be zero.
6689 // If so, we can compare against the unshifted value:
6690 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
6691 if (LHSI->hasOneUse() &&
6692 MaskedValueIsZero(LHSI->getOperand(0),
6693 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
6694 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6695 ConstantExpr::getShl(RHS, ShAmt));
6698 if (LHSI->hasOneUse()) {
6699 // Otherwise strength reduce the shift into an and.
6700 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
6701 Constant *Mask = ConstantInt::get(Val);
6704 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6705 Mask, LHSI->getName()+".mask");
6706 Value *And = InsertNewInstBefore(AndI, ICI);
6707 return new ICmpInst(ICI.getPredicate(), And,
6708 ConstantExpr::getShl(RHS, ShAmt));
6713 case Instruction::SDiv:
6714 case Instruction::UDiv:
6715 // Fold: icmp pred ([us]div X, C1), C2 -> range test
6716 // Fold this div into the comparison, producing a range check.
6717 // Determine, based on the divide type, what the range is being
6718 // checked. If there is an overflow on the low or high side, remember
6719 // it, otherwise compute the range [low, hi) bounding the new value.
6720 // See: InsertRangeTest above for the kinds of replacements possible.
6721 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
6722 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
6727 case Instruction::Add:
6728 // Fold: icmp pred (add, X, C1), C2
6730 if (!ICI.isEquality()) {
6731 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6733 const APInt &LHSV = LHSC->getValue();
6735 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
6738 if (ICI.isSignedPredicate()) {
6739 if (CR.getLower().isSignBit()) {
6740 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
6741 ConstantInt::get(CR.getUpper()));
6742 } else if (CR.getUpper().isSignBit()) {
6743 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
6744 ConstantInt::get(CR.getLower()));
6747 if (CR.getLower().isMinValue()) {
6748 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
6749 ConstantInt::get(CR.getUpper()));
6750 } else if (CR.getUpper().isMinValue()) {
6751 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
6752 ConstantInt::get(CR.getLower()));
6759 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
6760 if (ICI.isEquality()) {
6761 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6763 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
6764 // the second operand is a constant, simplify a bit.
6765 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
6766 switch (BO->getOpcode()) {
6767 case Instruction::SRem:
6768 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
6769 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
6770 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
6771 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
6772 Instruction *NewRem =
6773 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
6775 InsertNewInstBefore(NewRem, ICI);
6776 return new ICmpInst(ICI.getPredicate(), NewRem,
6777 Constant::getNullValue(BO->getType()));
6781 case Instruction::Add:
6782 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
6783 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6784 if (BO->hasOneUse())
6785 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6786 Subtract(RHS, BOp1C));
6787 } else if (RHSV == 0) {
6788 // Replace ((add A, B) != 0) with (A != -B) if A or B is
6789 // efficiently invertible, or if the add has just this one use.
6790 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
6792 if (Value *NegVal = dyn_castNegVal(BOp1))
6793 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
6794 else if (Value *NegVal = dyn_castNegVal(BOp0))
6795 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
6796 else if (BO->hasOneUse()) {
6797 Instruction *Neg = BinaryOperator::CreateNeg(BOp1);
6798 InsertNewInstBefore(Neg, ICI);
6800 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
6804 case Instruction::Xor:
6805 // For the xor case, we can xor two constants together, eliminating
6806 // the explicit xor.
6807 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
6808 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6809 ConstantExpr::getXor(RHS, BOC));
6812 case Instruction::Sub:
6813 // Replace (([sub|xor] A, B) != 0) with (A != B)
6815 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6819 case Instruction::Or:
6820 // If bits are being or'd in that are not present in the constant we
6821 // are comparing against, then the comparison could never succeed!
6822 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
6823 Constant *NotCI = ConstantExpr::getNot(RHS);
6824 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
6825 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6830 case Instruction::And:
6831 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6832 // If bits are being compared against that are and'd out, then the
6833 // comparison can never succeed!
6834 if ((RHSV & ~BOC->getValue()) != 0)
6835 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6838 // If we have ((X & C) == C), turn it into ((X & C) != 0).
6839 if (RHS == BOC && RHSV.isPowerOf2())
6840 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
6841 ICmpInst::ICMP_NE, LHSI,
6842 Constant::getNullValue(RHS->getType()));
6844 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
6845 if (BOC->getValue().isSignBit()) {
6846 Value *X = BO->getOperand(0);
6847 Constant *Zero = Constant::getNullValue(X->getType());
6848 ICmpInst::Predicate pred = isICMP_NE ?
6849 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
6850 return new ICmpInst(pred, X, Zero);
6853 // ((X & ~7) == 0) --> X < 8
6854 if (RHSV == 0 && isHighOnes(BOC)) {
6855 Value *X = BO->getOperand(0);
6856 Constant *NegX = ConstantExpr::getNeg(BOC);
6857 ICmpInst::Predicate pred = isICMP_NE ?
6858 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
6859 return new ICmpInst(pred, X, NegX);
6864 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
6865 // Handle icmp {eq|ne} <intrinsic>, intcst.
6866 if (II->getIntrinsicID() == Intrinsic::bswap) {
6868 ICI.setOperand(0, II->getOperand(1));
6869 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
6877 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
6878 /// We only handle extending casts so far.
6880 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
6881 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
6882 Value *LHSCIOp = LHSCI->getOperand(0);
6883 const Type *SrcTy = LHSCIOp->getType();
6884 const Type *DestTy = LHSCI->getType();
6887 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
6888 // integer type is the same size as the pointer type.
6889 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
6890 getTargetData().getPointerSizeInBits() ==
6891 cast<IntegerType>(DestTy)->getBitWidth()) {
6893 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
6894 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
6895 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
6896 RHSOp = RHSC->getOperand(0);
6897 // If the pointer types don't match, insert a bitcast.
6898 if (LHSCIOp->getType() != RHSOp->getType())
6899 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
6903 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
6906 // The code below only handles extension cast instructions, so far.
6908 if (LHSCI->getOpcode() != Instruction::ZExt &&
6909 LHSCI->getOpcode() != Instruction::SExt)
6912 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
6913 bool isSignedCmp = ICI.isSignedPredicate();
6915 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
6916 // Not an extension from the same type?
6917 RHSCIOp = CI->getOperand(0);
6918 if (RHSCIOp->getType() != LHSCIOp->getType())
6921 // If the signedness of the two casts doesn't agree (i.e. one is a sext
6922 // and the other is a zext), then we can't handle this.
6923 if (CI->getOpcode() != LHSCI->getOpcode())
6926 // Deal with equality cases early.
6927 if (ICI.isEquality())
6928 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6930 // A signed comparison of sign extended values simplifies into a
6931 // signed comparison.
6932 if (isSignedCmp && isSignedExt)
6933 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6935 // The other three cases all fold into an unsigned comparison.
6936 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
6939 // If we aren't dealing with a constant on the RHS, exit early
6940 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
6944 // Compute the constant that would happen if we truncated to SrcTy then
6945 // reextended to DestTy.
6946 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
6947 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
6949 // If the re-extended constant didn't change...
6951 // Make sure that sign of the Cmp and the sign of the Cast are the same.
6952 // For example, we might have:
6953 // %A = sext short %X to uint
6954 // %B = icmp ugt uint %A, 1330
6955 // It is incorrect to transform this into
6956 // %B = icmp ugt short %X, 1330
6957 // because %A may have negative value.
6959 // However, we allow this when the compare is EQ/NE, because they are
6961 if (isSignedExt == isSignedCmp || ICI.isEquality())
6962 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
6966 // The re-extended constant changed so the constant cannot be represented
6967 // in the shorter type. Consequently, we cannot emit a simple comparison.
6969 // First, handle some easy cases. We know the result cannot be equal at this
6970 // point so handle the ICI.isEquality() cases
6971 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6972 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6973 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6974 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6976 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
6977 // should have been folded away previously and not enter in here.
6980 // We're performing a signed comparison.
6981 if (cast<ConstantInt>(CI)->getValue().isNegative())
6982 Result = ConstantInt::getFalse(); // X < (small) --> false
6984 Result = ConstantInt::getTrue(); // X < (large) --> true
6986 // We're performing an unsigned comparison.
6988 // We're performing an unsigned comp with a sign extended value.
6989 // This is true if the input is >= 0. [aka >s -1]
6990 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
6991 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
6992 NegOne, ICI.getName()), ICI);
6994 // Unsigned extend & unsigned compare -> always true.
6995 Result = ConstantInt::getTrue();
6999 // Finally, return the value computed.
7000 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
7001 ICI.getPredicate() == ICmpInst::ICMP_SLT)
7002 return ReplaceInstUsesWith(ICI, Result);
7004 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
7005 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
7006 "ICmp should be folded!");
7007 if (Constant *CI = dyn_cast<Constant>(Result))
7008 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
7009 return BinaryOperator::CreateNot(Result);
7012 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
7013 return commonShiftTransforms(I);
7016 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
7017 return commonShiftTransforms(I);
7020 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
7021 if (Instruction *R = commonShiftTransforms(I))
7024 Value *Op0 = I.getOperand(0);
7026 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
7027 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
7028 if (CSI->isAllOnesValue())
7029 return ReplaceInstUsesWith(I, CSI);
7031 // See if we can turn a signed shr into an unsigned shr.
7032 if (!isa<VectorType>(I.getType()) &&
7033 MaskedValueIsZero(Op0,
7034 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
7035 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
7037 // Arithmetic shifting an all-sign-bit value is a no-op.
7038 unsigned NumSignBits = ComputeNumSignBits(Op0);
7039 if (NumSignBits == Op0->getType()->getPrimitiveSizeInBits())
7040 return ReplaceInstUsesWith(I, Op0);
7045 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
7046 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
7047 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
7049 // shl X, 0 == X and shr X, 0 == X
7050 // shl 0, X == 0 and shr 0, X == 0
7051 if (Op1 == Constant::getNullValue(Op1->getType()) ||
7052 Op0 == Constant::getNullValue(Op0->getType()))
7053 return ReplaceInstUsesWith(I, Op0);
7055 if (isa<UndefValue>(Op0)) {
7056 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
7057 return ReplaceInstUsesWith(I, Op0);
7058 else // undef << X -> 0, undef >>u X -> 0
7059 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7061 if (isa<UndefValue>(Op1)) {
7062 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
7063 return ReplaceInstUsesWith(I, Op0);
7064 else // X << undef, X >>u undef -> 0
7065 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7068 // Try to fold constant and into select arguments.
7069 if (isa<Constant>(Op0))
7070 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
7071 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7074 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
7075 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
7080 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
7081 BinaryOperator &I) {
7082 bool isLeftShift = I.getOpcode() == Instruction::Shl;
7084 // See if we can simplify any instructions used by the instruction whose sole
7085 // purpose is to compute bits we don't care about.
7086 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
7087 if (SimplifyDemandedInstructionBits(I))
7090 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
7091 // of a signed value.
7093 if (Op1->uge(TypeBits)) {
7094 if (I.getOpcode() != Instruction::AShr)
7095 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
7097 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
7102 // ((X*C1) << C2) == (X * (C1 << C2))
7103 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
7104 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
7105 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
7106 return BinaryOperator::CreateMul(BO->getOperand(0),
7107 ConstantExpr::getShl(BOOp, Op1));
7109 // Try to fold constant and into select arguments.
7110 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
7111 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7113 if (isa<PHINode>(Op0))
7114 if (Instruction *NV = FoldOpIntoPhi(I))
7117 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
7118 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
7119 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
7120 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
7121 // place. Don't try to do this transformation in this case. Also, we
7122 // require that the input operand is a shift-by-constant so that we have
7123 // confidence that the shifts will get folded together. We could do this
7124 // xform in more cases, but it is unlikely to be profitable.
7125 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
7126 isa<ConstantInt>(TrOp->getOperand(1))) {
7127 // Okay, we'll do this xform. Make the shift of shift.
7128 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
7129 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
7131 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
7133 // For logical shifts, the truncation has the effect of making the high
7134 // part of the register be zeros. Emulate this by inserting an AND to
7135 // clear the top bits as needed. This 'and' will usually be zapped by
7136 // other xforms later if dead.
7137 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
7138 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
7139 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
7141 // The mask we constructed says what the trunc would do if occurring
7142 // between the shifts. We want to know the effect *after* the second
7143 // shift. We know that it is a logical shift by a constant, so adjust the
7144 // mask as appropriate.
7145 if (I.getOpcode() == Instruction::Shl)
7146 MaskV <<= Op1->getZExtValue();
7148 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
7149 MaskV = MaskV.lshr(Op1->getZExtValue());
7152 Instruction *And = BinaryOperator::CreateAnd(NSh, ConstantInt::get(MaskV),
7154 InsertNewInstBefore(And, I); // shift1 & 0x00FF
7156 // Return the value truncated to the interesting size.
7157 return new TruncInst(And, I.getType());
7161 if (Op0->hasOneUse()) {
7162 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
7163 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7166 switch (Op0BO->getOpcode()) {
7168 case Instruction::Add:
7169 case Instruction::And:
7170 case Instruction::Or:
7171 case Instruction::Xor: {
7172 // These operators commute.
7173 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
7174 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
7175 match(Op0BO->getOperand(1), m_Shr(m_Value(V1), m_Specific(Op1)))){
7176 Instruction *YS = BinaryOperator::CreateShl(
7177 Op0BO->getOperand(0), Op1,
7179 InsertNewInstBefore(YS, I); // (Y << C)
7181 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
7182 Op0BO->getOperand(1)->getName());
7183 InsertNewInstBefore(X, I); // (X + (Y << C))
7184 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7185 return BinaryOperator::CreateAnd(X, ConstantInt::get(
7186 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7189 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
7190 Value *Op0BOOp1 = Op0BO->getOperand(1);
7191 if (isLeftShift && Op0BOOp1->hasOneUse() &&
7193 m_And(m_Shr(m_Value(V1), m_Specific(Op1)),
7194 m_ConstantInt(CC))) &&
7195 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
7196 Instruction *YS = BinaryOperator::CreateShl(
7197 Op0BO->getOperand(0), Op1,
7199 InsertNewInstBefore(YS, I); // (Y << C)
7201 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7202 V1->getName()+".mask");
7203 InsertNewInstBefore(XM, I); // X & (CC << C)
7205 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
7210 case Instruction::Sub: {
7211 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7212 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7213 match(Op0BO->getOperand(0), m_Shr(m_Value(V1), m_Specific(Op1)))){
7214 Instruction *YS = BinaryOperator::CreateShl(
7215 Op0BO->getOperand(1), Op1,
7217 InsertNewInstBefore(YS, I); // (Y << C)
7219 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
7220 Op0BO->getOperand(0)->getName());
7221 InsertNewInstBefore(X, I); // (X + (Y << C))
7222 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7223 return BinaryOperator::CreateAnd(X, ConstantInt::get(
7224 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7227 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7228 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7229 match(Op0BO->getOperand(0),
7230 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7231 m_ConstantInt(CC))) && V2 == Op1 &&
7232 cast<BinaryOperator>(Op0BO->getOperand(0))
7233 ->getOperand(0)->hasOneUse()) {
7234 Instruction *YS = BinaryOperator::CreateShl(
7235 Op0BO->getOperand(1), Op1,
7237 InsertNewInstBefore(YS, I); // (Y << C)
7239 BinaryOperator::CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7240 V1->getName()+".mask");
7241 InsertNewInstBefore(XM, I); // X & (CC << C)
7243 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7251 // If the operand is an bitwise operator with a constant RHS, and the
7252 // shift is the only use, we can pull it out of the shift.
7253 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7254 bool isValid = true; // Valid only for And, Or, Xor
7255 bool highBitSet = false; // Transform if high bit of constant set?
7257 switch (Op0BO->getOpcode()) {
7258 default: isValid = false; break; // Do not perform transform!
7259 case Instruction::Add:
7260 isValid = isLeftShift;
7262 case Instruction::Or:
7263 case Instruction::Xor:
7266 case Instruction::And:
7271 // If this is a signed shift right, and the high bit is modified
7272 // by the logical operation, do not perform the transformation.
7273 // The highBitSet boolean indicates the value of the high bit of
7274 // the constant which would cause it to be modified for this
7277 if (isValid && I.getOpcode() == Instruction::AShr)
7278 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7281 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
7283 Instruction *NewShift =
7284 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
7285 InsertNewInstBefore(NewShift, I);
7286 NewShift->takeName(Op0BO);
7288 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7295 // Find out if this is a shift of a shift by a constant.
7296 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7297 if (ShiftOp && !ShiftOp->isShift())
7300 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7301 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7302 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7303 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7304 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7305 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
7306 Value *X = ShiftOp->getOperand(0);
7308 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
7309 if (AmtSum > TypeBits)
7312 const IntegerType *Ty = cast<IntegerType>(I.getType());
7314 // Check for (X << c1) << c2 and (X >> c1) >> c2
7315 if (I.getOpcode() == ShiftOp->getOpcode()) {
7316 return BinaryOperator::Create(I.getOpcode(), X,
7317 ConstantInt::get(Ty, AmtSum));
7318 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
7319 I.getOpcode() == Instruction::AShr) {
7320 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
7321 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
7322 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
7323 I.getOpcode() == Instruction::LShr) {
7324 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
7325 Instruction *Shift =
7326 BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
7327 InsertNewInstBefore(Shift, I);
7329 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7330 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7333 // Okay, if we get here, one shift must be left, and the other shift must be
7334 // right. See if the amounts are equal.
7335 if (ShiftAmt1 == ShiftAmt2) {
7336 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
7337 if (I.getOpcode() == Instruction::Shl) {
7338 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
7339 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
7341 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
7342 if (I.getOpcode() == Instruction::LShr) {
7343 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
7344 return BinaryOperator::CreateAnd(X, ConstantInt::get(Mask));
7346 // We can simplify ((X << C) >>s C) into a trunc + sext.
7347 // NOTE: we could do this for any C, but that would make 'unusual' integer
7348 // types. For now, just stick to ones well-supported by the code
7350 const Type *SExtType = 0;
7351 switch (Ty->getBitWidth() - ShiftAmt1) {
7358 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
7363 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
7364 InsertNewInstBefore(NewTrunc, I);
7365 return new SExtInst(NewTrunc, Ty);
7367 // Otherwise, we can't handle it yet.
7368 } else if (ShiftAmt1 < ShiftAmt2) {
7369 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
7371 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
7372 if (I.getOpcode() == Instruction::Shl) {
7373 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7374 ShiftOp->getOpcode() == Instruction::AShr);
7375 Instruction *Shift =
7376 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7377 InsertNewInstBefore(Shift, I);
7379 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7380 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7383 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
7384 if (I.getOpcode() == Instruction::LShr) {
7385 assert(ShiftOp->getOpcode() == Instruction::Shl);
7386 Instruction *Shift =
7387 BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
7388 InsertNewInstBefore(Shift, I);
7390 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7391 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7394 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
7396 assert(ShiftAmt2 < ShiftAmt1);
7397 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
7399 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
7400 if (I.getOpcode() == Instruction::Shl) {
7401 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7402 ShiftOp->getOpcode() == Instruction::AShr);
7403 Instruction *Shift =
7404 BinaryOperator::Create(ShiftOp->getOpcode(), X,
7405 ConstantInt::get(Ty, ShiftDiff));
7406 InsertNewInstBefore(Shift, I);
7408 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7409 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7412 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
7413 if (I.getOpcode() == Instruction::LShr) {
7414 assert(ShiftOp->getOpcode() == Instruction::Shl);
7415 Instruction *Shift =
7416 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7417 InsertNewInstBefore(Shift, I);
7419 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7420 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(Mask));
7423 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
7430 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
7431 /// expression. If so, decompose it, returning some value X, such that Val is
7434 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
7436 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
7437 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
7438 Offset = CI->getZExtValue();
7440 return ConstantInt::get(Type::Int32Ty, 0);
7441 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
7442 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7443 if (I->getOpcode() == Instruction::Shl) {
7444 // This is a value scaled by '1 << the shift amt'.
7445 Scale = 1U << RHS->getZExtValue();
7447 return I->getOperand(0);
7448 } else if (I->getOpcode() == Instruction::Mul) {
7449 // This value is scaled by 'RHS'.
7450 Scale = RHS->getZExtValue();
7452 return I->getOperand(0);
7453 } else if (I->getOpcode() == Instruction::Add) {
7454 // We have X+C. Check to see if we really have (X*C2)+C1,
7455 // where C1 is divisible by C2.
7458 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
7459 Offset += RHS->getZExtValue();
7466 // Otherwise, we can't look past this.
7473 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
7474 /// try to eliminate the cast by moving the type information into the alloc.
7475 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
7476 AllocationInst &AI) {
7477 const PointerType *PTy = cast<PointerType>(CI.getType());
7479 // Remove any uses of AI that are dead.
7480 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
7482 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
7483 Instruction *User = cast<Instruction>(*UI++);
7484 if (isInstructionTriviallyDead(User)) {
7485 while (UI != E && *UI == User)
7486 ++UI; // If this instruction uses AI more than once, don't break UI.
7489 DOUT << "IC: DCE: " << *User;
7490 EraseInstFromFunction(*User);
7494 // Get the type really allocated and the type casted to.
7495 const Type *AllocElTy = AI.getAllocatedType();
7496 const Type *CastElTy = PTy->getElementType();
7497 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
7499 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
7500 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
7501 if (CastElTyAlign < AllocElTyAlign) return 0;
7503 // If the allocation has multiple uses, only promote it if we are strictly
7504 // increasing the alignment of the resultant allocation. If we keep it the
7505 // same, we open the door to infinite loops of various kinds. (A reference
7506 // from a dbg.declare doesn't count as a use for this purpose.)
7507 if (!AI.hasOneUse() && !hasOneUsePlusDeclare(&AI) &&
7508 CastElTyAlign == AllocElTyAlign) return 0;
7510 uint64_t AllocElTySize = TD->getTypePaddedSize(AllocElTy);
7511 uint64_t CastElTySize = TD->getTypePaddedSize(CastElTy);
7512 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
7514 // See if we can satisfy the modulus by pulling a scale out of the array
7516 unsigned ArraySizeScale;
7518 Value *NumElements = // See if the array size is a decomposable linear expr.
7519 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
7521 // If we can now satisfy the modulus, by using a non-1 scale, we really can
7523 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
7524 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
7526 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
7531 // If the allocation size is constant, form a constant mul expression
7532 Amt = ConstantInt::get(Type::Int32Ty, Scale);
7533 if (isa<ConstantInt>(NumElements))
7534 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
7535 // otherwise multiply the amount and the number of elements
7536 else if (Scale != 1) {
7537 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
7538 Amt = InsertNewInstBefore(Tmp, AI);
7542 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
7543 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
7544 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
7545 Amt = InsertNewInstBefore(Tmp, AI);
7548 AllocationInst *New;
7549 if (isa<MallocInst>(AI))
7550 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
7552 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
7553 InsertNewInstBefore(New, AI);
7556 // If the allocation has one real use plus a dbg.declare, just remove the
7558 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(&AI)) {
7559 EraseInstFromFunction(*DI);
7561 // If the allocation has multiple real uses, insert a cast and change all
7562 // things that used it to use the new cast. This will also hack on CI, but it
7564 else if (!AI.hasOneUse()) {
7565 AddUsesToWorkList(AI);
7566 // New is the allocation instruction, pointer typed. AI is the original
7567 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7568 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
7569 InsertNewInstBefore(NewCast, AI);
7570 AI.replaceAllUsesWith(NewCast);
7572 return ReplaceInstUsesWith(CI, New);
7575 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7576 /// and return it as type Ty without inserting any new casts and without
7577 /// changing the computed value. This is used by code that tries to decide
7578 /// whether promoting or shrinking integer operations to wider or smaller types
7579 /// will allow us to eliminate a truncate or extend.
7581 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7582 /// extension operation if Ty is larger.
7584 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
7585 /// should return true if trunc(V) can be computed by computing V in the smaller
7586 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
7587 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
7588 /// efficiently truncated.
7590 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
7591 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
7592 /// the final result.
7593 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
7595 int &NumCastsRemoved){
7596 // We can always evaluate constants in another type.
7597 if (isa<ConstantInt>(V))
7600 Instruction *I = dyn_cast<Instruction>(V);
7601 if (!I) return false;
7603 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
7605 // If this is an extension or truncate, we can often eliminate it.
7606 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7607 // If this is a cast from the destination type, we can trivially eliminate
7608 // it, and this will remove a cast overall.
7609 if (I->getOperand(0)->getType() == Ty) {
7610 // If the first operand is itself a cast, and is eliminable, do not count
7611 // this as an eliminable cast. We would prefer to eliminate those two
7613 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
7619 // We can't extend or shrink something that has multiple uses: doing so would
7620 // require duplicating the instruction in general, which isn't profitable.
7621 if (!I->hasOneUse()) return false;
7623 unsigned Opc = I->getOpcode();
7625 case Instruction::Add:
7626 case Instruction::Sub:
7627 case Instruction::Mul:
7628 case Instruction::And:
7629 case Instruction::Or:
7630 case Instruction::Xor:
7631 // These operators can all arbitrarily be extended or truncated.
7632 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7634 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7637 case Instruction::Shl:
7638 // If we are truncating the result of this SHL, and if it's a shift of a
7639 // constant amount, we can always perform a SHL in a smaller type.
7640 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7641 uint32_t BitWidth = Ty->getBitWidth();
7642 if (BitWidth < OrigTy->getBitWidth() &&
7643 CI->getLimitedValue(BitWidth) < BitWidth)
7644 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7648 case Instruction::LShr:
7649 // If this is a truncate of a logical shr, we can truncate it to a smaller
7650 // lshr iff we know that the bits we would otherwise be shifting in are
7652 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7653 uint32_t OrigBitWidth = OrigTy->getBitWidth();
7654 uint32_t BitWidth = Ty->getBitWidth();
7655 if (BitWidth < OrigBitWidth &&
7656 MaskedValueIsZero(I->getOperand(0),
7657 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
7658 CI->getLimitedValue(BitWidth) < BitWidth) {
7659 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7664 case Instruction::ZExt:
7665 case Instruction::SExt:
7666 case Instruction::Trunc:
7667 // If this is the same kind of case as our original (e.g. zext+zext), we
7668 // can safely replace it. Note that replacing it does not reduce the number
7669 // of casts in the input.
7673 // sext (zext ty1), ty2 -> zext ty2
7674 if (CastOpc == Instruction::SExt && Opc == Instruction::ZExt)
7677 case Instruction::Select: {
7678 SelectInst *SI = cast<SelectInst>(I);
7679 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
7681 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
7684 case Instruction::PHI: {
7685 // We can change a phi if we can change all operands.
7686 PHINode *PN = cast<PHINode>(I);
7687 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
7688 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
7694 // TODO: Can handle more cases here.
7701 /// EvaluateInDifferentType - Given an expression that
7702 /// CanEvaluateInDifferentType returns true for, actually insert the code to
7703 /// evaluate the expression.
7704 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
7706 if (Constant *C = dyn_cast<Constant>(V))
7707 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
7709 // Otherwise, it must be an instruction.
7710 Instruction *I = cast<Instruction>(V);
7711 Instruction *Res = 0;
7712 unsigned Opc = I->getOpcode();
7714 case Instruction::Add:
7715 case Instruction::Sub:
7716 case Instruction::Mul:
7717 case Instruction::And:
7718 case Instruction::Or:
7719 case Instruction::Xor:
7720 case Instruction::AShr:
7721 case Instruction::LShr:
7722 case Instruction::Shl: {
7723 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
7724 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7725 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
7728 case Instruction::Trunc:
7729 case Instruction::ZExt:
7730 case Instruction::SExt:
7731 // If the source type of the cast is the type we're trying for then we can
7732 // just return the source. There's no need to insert it because it is not
7734 if (I->getOperand(0)->getType() == Ty)
7735 return I->getOperand(0);
7737 // Otherwise, must be the same type of cast, so just reinsert a new one.
7738 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
7741 case Instruction::Select: {
7742 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7743 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
7744 Res = SelectInst::Create(I->getOperand(0), True, False);
7747 case Instruction::PHI: {
7748 PHINode *OPN = cast<PHINode>(I);
7749 PHINode *NPN = PHINode::Create(Ty);
7750 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
7751 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
7752 NPN->addIncoming(V, OPN->getIncomingBlock(i));
7758 // TODO: Can handle more cases here.
7759 assert(0 && "Unreachable!");
7764 return InsertNewInstBefore(Res, *I);
7767 /// @brief Implement the transforms common to all CastInst visitors.
7768 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
7769 Value *Src = CI.getOperand(0);
7771 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
7772 // eliminate it now.
7773 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7774 if (Instruction::CastOps opc =
7775 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
7776 // The first cast (CSrc) is eliminable so we need to fix up or replace
7777 // the second cast (CI). CSrc will then have a good chance of being dead.
7778 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
7782 // If we are casting a select then fold the cast into the select
7783 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
7784 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
7787 // If we are casting a PHI then fold the cast into the PHI
7788 if (isa<PHINode>(Src))
7789 if (Instruction *NV = FoldOpIntoPhi(CI))
7795 /// FindElementAtOffset - Given a type and a constant offset, determine whether
7796 /// or not there is a sequence of GEP indices into the type that will land us at
7797 /// the specified offset. If so, fill them into NewIndices and return the
7798 /// resultant element type, otherwise return null.
7799 static const Type *FindElementAtOffset(const Type *Ty, int64_t Offset,
7800 SmallVectorImpl<Value*> &NewIndices,
7801 const TargetData *TD) {
7802 if (!Ty->isSized()) return 0;
7804 // Start with the index over the outer type. Note that the type size
7805 // might be zero (even if the offset isn't zero) if the indexed type
7806 // is something like [0 x {int, int}]
7807 const Type *IntPtrTy = TD->getIntPtrType();
7808 int64_t FirstIdx = 0;
7809 if (int64_t TySize = TD->getTypePaddedSize(Ty)) {
7810 FirstIdx = Offset/TySize;
7811 Offset -= FirstIdx*TySize;
7813 // Handle hosts where % returns negative instead of values [0..TySize).
7817 assert(Offset >= 0);
7819 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
7822 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
7824 // Index into the types. If we fail, set OrigBase to null.
7826 // Indexing into tail padding between struct/array elements.
7827 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
7830 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
7831 const StructLayout *SL = TD->getStructLayout(STy);
7832 assert(Offset < (int64_t)SL->getSizeInBytes() &&
7833 "Offset must stay within the indexed type");
7835 unsigned Elt = SL->getElementContainingOffset(Offset);
7836 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
7838 Offset -= SL->getElementOffset(Elt);
7839 Ty = STy->getElementType(Elt);
7840 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
7841 uint64_t EltSize = TD->getTypePaddedSize(AT->getElementType());
7842 assert(EltSize && "Cannot index into a zero-sized array");
7843 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
7845 Ty = AT->getElementType();
7847 // Otherwise, we can't index into the middle of this atomic type, bail.
7855 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
7856 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
7857 Value *Src = CI.getOperand(0);
7859 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
7860 // If casting the result of a getelementptr instruction with no offset, turn
7861 // this into a cast of the original pointer!
7862 if (GEP->hasAllZeroIndices()) {
7863 // Changing the cast operand is usually not a good idea but it is safe
7864 // here because the pointer operand is being replaced with another
7865 // pointer operand so the opcode doesn't need to change.
7867 CI.setOperand(0, GEP->getOperand(0));
7871 // If the GEP has a single use, and the base pointer is a bitcast, and the
7872 // GEP computes a constant offset, see if we can convert these three
7873 // instructions into fewer. This typically happens with unions and other
7874 // non-type-safe code.
7875 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
7876 if (GEP->hasAllConstantIndices()) {
7877 // We are guaranteed to get a constant from EmitGEPOffset.
7878 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
7879 int64_t Offset = OffsetV->getSExtValue();
7881 // Get the base pointer input of the bitcast, and the type it points to.
7882 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
7883 const Type *GEPIdxTy =
7884 cast<PointerType>(OrigBase->getType())->getElementType();
7885 SmallVector<Value*, 8> NewIndices;
7886 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices, TD)) {
7887 // If we were able to index down into an element, create the GEP
7888 // and bitcast the result. This eliminates one bitcast, potentially
7890 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
7892 NewIndices.end(), "");
7893 InsertNewInstBefore(NGEP, CI);
7894 NGEP->takeName(GEP);
7896 if (isa<BitCastInst>(CI))
7897 return new BitCastInst(NGEP, CI.getType());
7898 assert(isa<PtrToIntInst>(CI));
7899 return new PtrToIntInst(NGEP, CI.getType());
7905 return commonCastTransforms(CI);
7909 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
7910 /// integer types. This function implements the common transforms for all those
7912 /// @brief Implement the transforms common to CastInst with integer operands
7913 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
7914 if (Instruction *Result = commonCastTransforms(CI))
7917 Value *Src = CI.getOperand(0);
7918 const Type *SrcTy = Src->getType();
7919 const Type *DestTy = CI.getType();
7920 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
7921 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
7923 // See if we can simplify any instructions used by the LHS whose sole
7924 // purpose is to compute bits we don't care about.
7925 if (SimplifyDemandedInstructionBits(CI))
7928 // If the source isn't an instruction or has more than one use then we
7929 // can't do anything more.
7930 Instruction *SrcI = dyn_cast<Instruction>(Src);
7931 if (!SrcI || !Src->hasOneUse())
7934 // Attempt to propagate the cast into the instruction for int->int casts.
7935 int NumCastsRemoved = 0;
7936 if (!isa<BitCastInst>(CI) &&
7937 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
7938 CI.getOpcode(), NumCastsRemoved)) {
7939 // If this cast is a truncate, evaluting in a different type always
7940 // eliminates the cast, so it is always a win. If this is a zero-extension,
7941 // we need to do an AND to maintain the clear top-part of the computation,
7942 // so we require that the input have eliminated at least one cast. If this
7943 // is a sign extension, we insert two new casts (to do the extension) so we
7944 // require that two casts have been eliminated.
7945 bool DoXForm = false;
7946 bool JustReplace = false;
7947 switch (CI.getOpcode()) {
7949 // All the others use floating point so we shouldn't actually
7950 // get here because of the check above.
7951 assert(0 && "Unknown cast type");
7952 case Instruction::Trunc:
7955 case Instruction::ZExt: {
7956 DoXForm = NumCastsRemoved >= 1;
7957 if (!DoXForm && 0) {
7958 // If it's unnecessary to issue an AND to clear the high bits, it's
7959 // always profitable to do this xform.
7960 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, false);
7961 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
7962 if (MaskedValueIsZero(TryRes, Mask))
7963 return ReplaceInstUsesWith(CI, TryRes);
7965 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
7966 if (TryI->use_empty())
7967 EraseInstFromFunction(*TryI);
7971 case Instruction::SExt: {
7972 DoXForm = NumCastsRemoved >= 2;
7973 if (!DoXForm && !isa<TruncInst>(SrcI) && 0) {
7974 // If we do not have to emit the truncate + sext pair, then it's always
7975 // profitable to do this xform.
7977 // It's not safe to eliminate the trunc + sext pair if one of the
7978 // eliminated cast is a truncate. e.g.
7979 // t2 = trunc i32 t1 to i16
7980 // t3 = sext i16 t2 to i32
7983 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, true);
7984 unsigned NumSignBits = ComputeNumSignBits(TryRes);
7985 if (NumSignBits > (DestBitSize - SrcBitSize))
7986 return ReplaceInstUsesWith(CI, TryRes);
7988 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
7989 if (TryI->use_empty())
7990 EraseInstFromFunction(*TryI);
7997 DOUT << "ICE: EvaluateInDifferentType converting expression type to avoid"
7999 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
8000 CI.getOpcode() == Instruction::SExt);
8002 // Just replace this cast with the result.
8003 return ReplaceInstUsesWith(CI, Res);
8005 assert(Res->getType() == DestTy);
8006 switch (CI.getOpcode()) {
8007 default: assert(0 && "Unknown cast type!");
8008 case Instruction::Trunc:
8009 case Instruction::BitCast:
8010 // Just replace this cast with the result.
8011 return ReplaceInstUsesWith(CI, Res);
8012 case Instruction::ZExt: {
8013 assert(SrcBitSize < DestBitSize && "Not a zext?");
8015 // If the high bits are already zero, just replace this cast with the
8017 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8018 if (MaskedValueIsZero(Res, Mask))
8019 return ReplaceInstUsesWith(CI, Res);
8021 // We need to emit an AND to clear the high bits.
8022 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
8024 return BinaryOperator::CreateAnd(Res, C);
8026 case Instruction::SExt: {
8027 // If the high bits are already filled with sign bit, just replace this
8028 // cast with the result.
8029 unsigned NumSignBits = ComputeNumSignBits(Res);
8030 if (NumSignBits > (DestBitSize - SrcBitSize))
8031 return ReplaceInstUsesWith(CI, Res);
8033 // We need to emit a cast to truncate, then a cast to sext.
8034 return CastInst::Create(Instruction::SExt,
8035 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
8042 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
8043 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
8045 switch (SrcI->getOpcode()) {
8046 case Instruction::Add:
8047 case Instruction::Mul:
8048 case Instruction::And:
8049 case Instruction::Or:
8050 case Instruction::Xor:
8051 // If we are discarding information, rewrite.
8052 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
8053 // Don't insert two casts if they cannot be eliminated. We allow
8054 // two casts to be inserted if the sizes are the same. This could
8055 // only be converting signedness, which is a noop.
8056 if (DestBitSize == SrcBitSize ||
8057 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
8058 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
8059 Instruction::CastOps opcode = CI.getOpcode();
8060 Value *Op0c = InsertCastBefore(opcode, Op0, DestTy, *SrcI);
8061 Value *Op1c = InsertCastBefore(opcode, Op1, DestTy, *SrcI);
8062 return BinaryOperator::Create(
8063 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
8067 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
8068 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
8069 SrcI->getOpcode() == Instruction::Xor &&
8070 Op1 == ConstantInt::getTrue() &&
8071 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
8072 Value *New = InsertCastBefore(Instruction::ZExt, Op0, DestTy, CI);
8073 return BinaryOperator::CreateXor(New, ConstantInt::get(CI.getType(), 1));
8076 case Instruction::SDiv:
8077 case Instruction::UDiv:
8078 case Instruction::SRem:
8079 case Instruction::URem:
8080 // If we are just changing the sign, rewrite.
8081 if (DestBitSize == SrcBitSize) {
8082 // Don't insert two casts if they cannot be eliminated. We allow
8083 // two casts to be inserted if the sizes are the same. This could
8084 // only be converting signedness, which is a noop.
8085 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
8086 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
8087 Value *Op0c = InsertCastBefore(Instruction::BitCast,
8088 Op0, DestTy, *SrcI);
8089 Value *Op1c = InsertCastBefore(Instruction::BitCast,
8090 Op1, DestTy, *SrcI);
8091 return BinaryOperator::Create(
8092 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
8097 case Instruction::Shl:
8098 // Allow changing the sign of the source operand. Do not allow
8099 // changing the size of the shift, UNLESS the shift amount is a
8100 // constant. We must not change variable sized shifts to a smaller
8101 // size, because it is undefined to shift more bits out than exist
8103 if (DestBitSize == SrcBitSize ||
8104 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
8105 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
8106 Instruction::BitCast : Instruction::Trunc);
8107 Value *Op0c = InsertCastBefore(opcode, Op0, DestTy, *SrcI);
8108 Value *Op1c = InsertCastBefore(opcode, Op1, DestTy, *SrcI);
8109 return BinaryOperator::CreateShl(Op0c, Op1c);
8112 case Instruction::AShr:
8113 // If this is a signed shr, and if all bits shifted in are about to be
8114 // truncated off, turn it into an unsigned shr to allow greater
8116 if (DestBitSize < SrcBitSize &&
8117 isa<ConstantInt>(Op1)) {
8118 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
8119 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
8120 // Insert the new logical shift right.
8121 return BinaryOperator::CreateLShr(Op0, Op1);
8129 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
8130 if (Instruction *Result = commonIntCastTransforms(CI))
8133 Value *Src = CI.getOperand(0);
8134 const Type *Ty = CI.getType();
8135 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
8136 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
8138 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
8139 switch (SrcI->getOpcode()) {
8141 case Instruction::LShr:
8142 // We can shrink lshr to something smaller if we know the bits shifted in
8143 // are already zeros.
8144 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
8145 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
8147 // Get a mask for the bits shifting in.
8148 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
8149 Value* SrcIOp0 = SrcI->getOperand(0);
8150 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
8151 if (ShAmt >= DestBitWidth) // All zeros.
8152 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
8154 // Okay, we can shrink this. Truncate the input, then return a new
8156 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
8157 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
8159 return BinaryOperator::CreateLShr(V1, V2);
8161 } else { // This is a variable shr.
8163 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
8164 // more LLVM instructions, but allows '1 << Y' to be hoisted if
8165 // loop-invariant and CSE'd.
8166 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
8167 Value *One = ConstantInt::get(SrcI->getType(), 1);
8169 Value *V = InsertNewInstBefore(
8170 BinaryOperator::CreateShl(One, SrcI->getOperand(1),
8172 V = InsertNewInstBefore(BinaryOperator::CreateAnd(V,
8173 SrcI->getOperand(0),
8175 Value *Zero = Constant::getNullValue(V->getType());
8176 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
8186 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
8187 /// in order to eliminate the icmp.
8188 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
8190 // If we are just checking for a icmp eq of a single bit and zext'ing it
8191 // to an integer, then shift the bit to the appropriate place and then
8192 // cast to integer to avoid the comparison.
8193 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
8194 const APInt &Op1CV = Op1C->getValue();
8196 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
8197 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
8198 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8199 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
8200 if (!DoXform) return ICI;
8202 Value *In = ICI->getOperand(0);
8203 Value *Sh = ConstantInt::get(In->getType(),
8204 In->getType()->getPrimitiveSizeInBits()-1);
8205 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
8206 In->getName()+".lobit"),
8208 if (In->getType() != CI.getType())
8209 In = CastInst::CreateIntegerCast(In, CI.getType(),
8210 false/*ZExt*/, "tmp", &CI);
8212 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
8213 Constant *One = ConstantInt::get(In->getType(), 1);
8214 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
8215 In->getName()+".not"),
8219 return ReplaceInstUsesWith(CI, In);
8224 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
8225 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8226 // zext (X == 1) to i32 --> X iff X has only the low bit set.
8227 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
8228 // zext (X != 0) to i32 --> X iff X has only the low bit set.
8229 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
8230 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
8231 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8232 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
8233 // This only works for EQ and NE
8234 ICI->isEquality()) {
8235 // If Op1C some other power of two, convert:
8236 uint32_t BitWidth = Op1C->getType()->getBitWidth();
8237 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8238 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8239 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
8241 APInt KnownZeroMask(~KnownZero);
8242 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
8243 if (!DoXform) return ICI;
8245 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
8246 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
8247 // (X&4) == 2 --> false
8248 // (X&4) != 2 --> true
8249 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
8250 Res = ConstantExpr::getZExt(Res, CI.getType());
8251 return ReplaceInstUsesWith(CI, Res);
8254 uint32_t ShiftAmt = KnownZeroMask.logBase2();
8255 Value *In = ICI->getOperand(0);
8257 // Perform a logical shr by shiftamt.
8258 // Insert the shift to put the result in the low bit.
8259 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
8260 ConstantInt::get(In->getType(), ShiftAmt),
8261 In->getName()+".lobit"), CI);
8264 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
8265 Constant *One = ConstantInt::get(In->getType(), 1);
8266 In = BinaryOperator::CreateXor(In, One, "tmp");
8267 InsertNewInstBefore(cast<Instruction>(In), CI);
8270 if (CI.getType() == In->getType())
8271 return ReplaceInstUsesWith(CI, In);
8273 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
8281 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
8282 // If one of the common conversion will work ..
8283 if (Instruction *Result = commonIntCastTransforms(CI))
8286 Value *Src = CI.getOperand(0);
8288 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
8289 // types and if the sizes are just right we can convert this into a logical
8290 // 'and' which will be much cheaper than the pair of casts.
8291 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
8292 // Get the sizes of the types involved. We know that the intermediate type
8293 // will be smaller than A or C, but don't know the relation between A and C.
8294 Value *A = CSrc->getOperand(0);
8295 unsigned SrcSize = A->getType()->getPrimitiveSizeInBits();
8296 unsigned MidSize = CSrc->getType()->getPrimitiveSizeInBits();
8297 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
8298 // If we're actually extending zero bits, then if
8299 // SrcSize < DstSize: zext(a & mask)
8300 // SrcSize == DstSize: a & mask
8301 // SrcSize > DstSize: trunc(a) & mask
8302 if (SrcSize < DstSize) {
8303 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8304 Constant *AndConst = ConstantInt::get(AndValue);
8306 BinaryOperator::CreateAnd(A, AndConst, CSrc->getName()+".mask");
8307 InsertNewInstBefore(And, CI);
8308 return new ZExtInst(And, CI.getType());
8309 } else if (SrcSize == DstSize) {
8310 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8311 return BinaryOperator::CreateAnd(A, ConstantInt::get(AndValue));
8312 } else if (SrcSize > DstSize) {
8313 Instruction *Trunc = new TruncInst(A, CI.getType(), "tmp");
8314 InsertNewInstBefore(Trunc, CI);
8315 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
8316 return BinaryOperator::CreateAnd(Trunc, ConstantInt::get(AndValue));
8320 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
8321 return transformZExtICmp(ICI, CI);
8323 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
8324 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
8325 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
8326 // of the (zext icmp) will be transformed.
8327 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
8328 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
8329 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
8330 (transformZExtICmp(LHS, CI, false) ||
8331 transformZExtICmp(RHS, CI, false))) {
8332 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
8333 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
8334 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
8341 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
8342 if (Instruction *I = commonIntCastTransforms(CI))
8345 Value *Src = CI.getOperand(0);
8347 // Canonicalize sign-extend from i1 to a select.
8348 if (Src->getType() == Type::Int1Ty)
8349 return SelectInst::Create(Src,
8350 ConstantInt::getAllOnesValue(CI.getType()),
8351 Constant::getNullValue(CI.getType()));
8353 // See if the value being truncated is already sign extended. If so, just
8354 // eliminate the trunc/sext pair.
8355 if (getOpcode(Src) == Instruction::Trunc) {
8356 Value *Op = cast<User>(Src)->getOperand(0);
8357 unsigned OpBits = cast<IntegerType>(Op->getType())->getBitWidth();
8358 unsigned MidBits = cast<IntegerType>(Src->getType())->getBitWidth();
8359 unsigned DestBits = cast<IntegerType>(CI.getType())->getBitWidth();
8360 unsigned NumSignBits = ComputeNumSignBits(Op);
8362 if (OpBits == DestBits) {
8363 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
8364 // bits, it is already ready.
8365 if (NumSignBits > DestBits-MidBits)
8366 return ReplaceInstUsesWith(CI, Op);
8367 } else if (OpBits < DestBits) {
8368 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
8369 // bits, just sext from i32.
8370 if (NumSignBits > OpBits-MidBits)
8371 return new SExtInst(Op, CI.getType(), "tmp");
8373 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
8374 // bits, just truncate to i32.
8375 if (NumSignBits > OpBits-MidBits)
8376 return new TruncInst(Op, CI.getType(), "tmp");
8380 // If the input is a shl/ashr pair of a same constant, then this is a sign
8381 // extension from a smaller value. If we could trust arbitrary bitwidth
8382 // integers, we could turn this into a truncate to the smaller bit and then
8383 // use a sext for the whole extension. Since we don't, look deeper and check
8384 // for a truncate. If the source and dest are the same type, eliminate the
8385 // trunc and extend and just do shifts. For example, turn:
8386 // %a = trunc i32 %i to i8
8387 // %b = shl i8 %a, 6
8388 // %c = ashr i8 %b, 6
8389 // %d = sext i8 %c to i32
8391 // %a = shl i32 %i, 30
8392 // %d = ashr i32 %a, 30
8394 ConstantInt *BA = 0, *CA = 0;
8395 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
8396 m_ConstantInt(CA))) &&
8397 BA == CA && isa<TruncInst>(A)) {
8398 Value *I = cast<TruncInst>(A)->getOperand(0);
8399 if (I->getType() == CI.getType()) {
8400 unsigned MidSize = Src->getType()->getPrimitiveSizeInBits();
8401 unsigned SrcDstSize = CI.getType()->getPrimitiveSizeInBits();
8402 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
8403 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
8404 I = InsertNewInstBefore(BinaryOperator::CreateShl(I, ShAmtV,
8406 return BinaryOperator::CreateAShr(I, ShAmtV);
8413 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
8414 /// in the specified FP type without changing its value.
8415 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
8417 APFloat F = CFP->getValueAPF();
8418 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
8420 return ConstantFP::get(F);
8424 /// LookThroughFPExtensions - If this is an fp extension instruction, look
8425 /// through it until we get the source value.
8426 static Value *LookThroughFPExtensions(Value *V) {
8427 if (Instruction *I = dyn_cast<Instruction>(V))
8428 if (I->getOpcode() == Instruction::FPExt)
8429 return LookThroughFPExtensions(I->getOperand(0));
8431 // If this value is a constant, return the constant in the smallest FP type
8432 // that can accurately represent it. This allows us to turn
8433 // (float)((double)X+2.0) into x+2.0f.
8434 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
8435 if (CFP->getType() == Type::PPC_FP128Ty)
8436 return V; // No constant folding of this.
8437 // See if the value can be truncated to float and then reextended.
8438 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
8440 if (CFP->getType() == Type::DoubleTy)
8441 return V; // Won't shrink.
8442 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
8444 // Don't try to shrink to various long double types.
8450 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
8451 if (Instruction *I = commonCastTransforms(CI))
8454 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
8455 // smaller than the destination type, we can eliminate the truncate by doing
8456 // the add as the smaller type. This applies to add/sub/mul/div as well as
8457 // many builtins (sqrt, etc).
8458 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
8459 if (OpI && OpI->hasOneUse()) {
8460 switch (OpI->getOpcode()) {
8462 case Instruction::Add:
8463 case Instruction::Sub:
8464 case Instruction::Mul:
8465 case Instruction::FDiv:
8466 case Instruction::FRem:
8467 const Type *SrcTy = OpI->getType();
8468 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
8469 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
8470 if (LHSTrunc->getType() != SrcTy &&
8471 RHSTrunc->getType() != SrcTy) {
8472 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
8473 // If the source types were both smaller than the destination type of
8474 // the cast, do this xform.
8475 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
8476 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
8477 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
8479 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
8481 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
8490 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
8491 return commonCastTransforms(CI);
8494 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
8495 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8497 return commonCastTransforms(FI);
8499 // fptoui(uitofp(X)) --> X
8500 // fptoui(sitofp(X)) --> X
8501 // This is safe if the intermediate type has enough bits in its mantissa to
8502 // accurately represent all values of X. For example, do not do this with
8503 // i64->float->i64. This is also safe for sitofp case, because any negative
8504 // 'X' value would cause an undefined result for the fptoui.
8505 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8506 OpI->getOperand(0)->getType() == FI.getType() &&
8507 (int)FI.getType()->getPrimitiveSizeInBits() < /*extra bit for sign */
8508 OpI->getType()->getFPMantissaWidth())
8509 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8511 return commonCastTransforms(FI);
8514 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
8515 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8517 return commonCastTransforms(FI);
8519 // fptosi(sitofp(X)) --> X
8520 // fptosi(uitofp(X)) --> X
8521 // This is safe if the intermediate type has enough bits in its mantissa to
8522 // accurately represent all values of X. For example, do not do this with
8523 // i64->float->i64. This is also safe for sitofp case, because any negative
8524 // 'X' value would cause an undefined result for the fptoui.
8525 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8526 OpI->getOperand(0)->getType() == FI.getType() &&
8527 (int)FI.getType()->getPrimitiveSizeInBits() <=
8528 OpI->getType()->getFPMantissaWidth())
8529 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8531 return commonCastTransforms(FI);
8534 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
8535 return commonCastTransforms(CI);
8538 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
8539 return commonCastTransforms(CI);
8542 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
8543 return commonPointerCastTransforms(CI);
8546 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
8547 if (Instruction *I = commonCastTransforms(CI))
8550 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
8551 if (!DestPointee->isSized()) return 0;
8553 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
8556 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
8557 m_ConstantInt(Cst)))) {
8558 // If the source and destination operands have the same type, see if this
8559 // is a single-index GEP.
8560 if (X->getType() == CI.getType()) {
8561 // Get the size of the pointee type.
8562 uint64_t Size = TD->getTypePaddedSize(DestPointee);
8564 // Convert the constant to intptr type.
8565 APInt Offset = Cst->getValue();
8566 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8568 // If Offset is evenly divisible by Size, we can do this xform.
8569 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8570 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8571 return GetElementPtrInst::Create(X, ConstantInt::get(Offset));
8574 // TODO: Could handle other cases, e.g. where add is indexing into field of
8576 } else if (CI.getOperand(0)->hasOneUse() &&
8577 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
8578 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
8579 // "inttoptr+GEP" instead of "add+intptr".
8581 // Get the size of the pointee type.
8582 uint64_t Size = TD->getTypePaddedSize(DestPointee);
8584 // Convert the constant to intptr type.
8585 APInt Offset = Cst->getValue();
8586 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8588 // If Offset is evenly divisible by Size, we can do this xform.
8589 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8590 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8592 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
8594 return GetElementPtrInst::Create(P, ConstantInt::get(Offset), "tmp");
8600 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
8601 // If the operands are integer typed then apply the integer transforms,
8602 // otherwise just apply the common ones.
8603 Value *Src = CI.getOperand(0);
8604 const Type *SrcTy = Src->getType();
8605 const Type *DestTy = CI.getType();
8607 if (SrcTy->isInteger() && DestTy->isInteger()) {
8608 if (Instruction *Result = commonIntCastTransforms(CI))
8610 } else if (isa<PointerType>(SrcTy)) {
8611 if (Instruction *I = commonPointerCastTransforms(CI))
8614 if (Instruction *Result = commonCastTransforms(CI))
8619 // Get rid of casts from one type to the same type. These are useless and can
8620 // be replaced by the operand.
8621 if (DestTy == Src->getType())
8622 return ReplaceInstUsesWith(CI, Src);
8624 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
8625 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
8626 const Type *DstElTy = DstPTy->getElementType();
8627 const Type *SrcElTy = SrcPTy->getElementType();
8629 // If the address spaces don't match, don't eliminate the bitcast, which is
8630 // required for changing types.
8631 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
8634 // If we are casting a malloc or alloca to a pointer to a type of the same
8635 // size, rewrite the allocation instruction to allocate the "right" type.
8636 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
8637 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
8640 // If the source and destination are pointers, and this cast is equivalent
8641 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
8642 // This can enhance SROA and other transforms that want type-safe pointers.
8643 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
8644 unsigned NumZeros = 0;
8645 while (SrcElTy != DstElTy &&
8646 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
8647 SrcElTy->getNumContainedTypes() /* not "{}" */) {
8648 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
8652 // If we found a path from the src to dest, create the getelementptr now.
8653 if (SrcElTy == DstElTy) {
8654 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
8655 return GetElementPtrInst::Create(Src, Idxs.begin(), Idxs.end(), "",
8656 ((Instruction*) NULL));
8660 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
8661 if (SVI->hasOneUse()) {
8662 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
8663 // a bitconvert to a vector with the same # elts.
8664 if (isa<VectorType>(DestTy) &&
8665 cast<VectorType>(DestTy)->getNumElements() ==
8666 SVI->getType()->getNumElements() &&
8667 SVI->getType()->getNumElements() ==
8668 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
8670 // If either of the operands is a cast from CI.getType(), then
8671 // evaluating the shuffle in the casted destination's type will allow
8672 // us to eliminate at least one cast.
8673 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
8674 Tmp->getOperand(0)->getType() == DestTy) ||
8675 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
8676 Tmp->getOperand(0)->getType() == DestTy)) {
8677 Value *LHS = InsertCastBefore(Instruction::BitCast,
8678 SVI->getOperand(0), DestTy, CI);
8679 Value *RHS = InsertCastBefore(Instruction::BitCast,
8680 SVI->getOperand(1), DestTy, CI);
8681 // Return a new shuffle vector. Use the same element ID's, as we
8682 // know the vector types match #elts.
8683 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
8691 /// GetSelectFoldableOperands - We want to turn code that looks like this:
8693 /// %D = select %cond, %C, %A
8695 /// %C = select %cond, %B, 0
8698 /// Assuming that the specified instruction is an operand to the select, return
8699 /// a bitmask indicating which operands of this instruction are foldable if they
8700 /// equal the other incoming value of the select.
8702 static unsigned GetSelectFoldableOperands(Instruction *I) {
8703 switch (I->getOpcode()) {
8704 case Instruction::Add:
8705 case Instruction::Mul:
8706 case Instruction::And:
8707 case Instruction::Or:
8708 case Instruction::Xor:
8709 return 3; // Can fold through either operand.
8710 case Instruction::Sub: // Can only fold on the amount subtracted.
8711 case Instruction::Shl: // Can only fold on the shift amount.
8712 case Instruction::LShr:
8713 case Instruction::AShr:
8716 return 0; // Cannot fold
8720 /// GetSelectFoldableConstant - For the same transformation as the previous
8721 /// function, return the identity constant that goes into the select.
8722 static Constant *GetSelectFoldableConstant(Instruction *I) {
8723 switch (I->getOpcode()) {
8724 default: assert(0 && "This cannot happen!"); abort();
8725 case Instruction::Add:
8726 case Instruction::Sub:
8727 case Instruction::Or:
8728 case Instruction::Xor:
8729 case Instruction::Shl:
8730 case Instruction::LShr:
8731 case Instruction::AShr:
8732 return Constant::getNullValue(I->getType());
8733 case Instruction::And:
8734 return Constant::getAllOnesValue(I->getType());
8735 case Instruction::Mul:
8736 return ConstantInt::get(I->getType(), 1);
8740 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
8741 /// have the same opcode and only one use each. Try to simplify this.
8742 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
8744 if (TI->getNumOperands() == 1) {
8745 // If this is a non-volatile load or a cast from the same type,
8748 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
8751 return 0; // unknown unary op.
8754 // Fold this by inserting a select from the input values.
8755 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
8756 FI->getOperand(0), SI.getName()+".v");
8757 InsertNewInstBefore(NewSI, SI);
8758 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
8762 // Only handle binary operators here.
8763 if (!isa<BinaryOperator>(TI))
8766 // Figure out if the operations have any operands in common.
8767 Value *MatchOp, *OtherOpT, *OtherOpF;
8769 if (TI->getOperand(0) == FI->getOperand(0)) {
8770 MatchOp = TI->getOperand(0);
8771 OtherOpT = TI->getOperand(1);
8772 OtherOpF = FI->getOperand(1);
8773 MatchIsOpZero = true;
8774 } else if (TI->getOperand(1) == FI->getOperand(1)) {
8775 MatchOp = TI->getOperand(1);
8776 OtherOpT = TI->getOperand(0);
8777 OtherOpF = FI->getOperand(0);
8778 MatchIsOpZero = false;
8779 } else if (!TI->isCommutative()) {
8781 } else if (TI->getOperand(0) == FI->getOperand(1)) {
8782 MatchOp = TI->getOperand(0);
8783 OtherOpT = TI->getOperand(1);
8784 OtherOpF = FI->getOperand(0);
8785 MatchIsOpZero = true;
8786 } else if (TI->getOperand(1) == FI->getOperand(0)) {
8787 MatchOp = TI->getOperand(1);
8788 OtherOpT = TI->getOperand(0);
8789 OtherOpF = FI->getOperand(1);
8790 MatchIsOpZero = true;
8795 // If we reach here, they do have operations in common.
8796 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
8797 OtherOpF, SI.getName()+".v");
8798 InsertNewInstBefore(NewSI, SI);
8800 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
8802 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
8804 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
8806 assert(0 && "Shouldn't get here");
8810 /// visitSelectInstWithICmp - Visit a SelectInst that has an
8811 /// ICmpInst as its first operand.
8813 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
8815 bool Changed = false;
8816 ICmpInst::Predicate Pred = ICI->getPredicate();
8817 Value *CmpLHS = ICI->getOperand(0);
8818 Value *CmpRHS = ICI->getOperand(1);
8819 Value *TrueVal = SI.getTrueValue();
8820 Value *FalseVal = SI.getFalseValue();
8822 // Check cases where the comparison is with a constant that
8823 // can be adjusted to fit the min/max idiom. We may edit ICI in
8824 // place here, so make sure the select is the only user.
8825 if (ICI->hasOneUse())
8826 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
8829 case ICmpInst::ICMP_ULT:
8830 case ICmpInst::ICMP_SLT: {
8831 // X < MIN ? T : F --> F
8832 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
8833 return ReplaceInstUsesWith(SI, FalseVal);
8834 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
8835 Constant *AdjustedRHS = SubOne(CI);
8836 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
8837 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
8838 Pred = ICmpInst::getSwappedPredicate(Pred);
8839 CmpRHS = AdjustedRHS;
8840 std::swap(FalseVal, TrueVal);
8841 ICI->setPredicate(Pred);
8842 ICI->setOperand(1, CmpRHS);
8843 SI.setOperand(1, TrueVal);
8844 SI.setOperand(2, FalseVal);
8849 case ICmpInst::ICMP_UGT:
8850 case ICmpInst::ICMP_SGT: {
8851 // X > MAX ? T : F --> F
8852 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
8853 return ReplaceInstUsesWith(SI, FalseVal);
8854 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
8855 Constant *AdjustedRHS = AddOne(CI);
8856 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
8857 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
8858 Pred = ICmpInst::getSwappedPredicate(Pred);
8859 CmpRHS = AdjustedRHS;
8860 std::swap(FalseVal, TrueVal);
8861 ICI->setPredicate(Pred);
8862 ICI->setOperand(1, CmpRHS);
8863 SI.setOperand(1, TrueVal);
8864 SI.setOperand(2, FalseVal);
8871 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
8872 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
8873 CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
8874 if (match(TrueVal, m_ConstantInt<-1>()) &&
8875 match(FalseVal, m_ConstantInt<0>()))
8876 Pred = ICI->getPredicate();
8877 else if (match(TrueVal, m_ConstantInt<0>()) &&
8878 match(FalseVal, m_ConstantInt<-1>()))
8879 Pred = CmpInst::getInversePredicate(ICI->getPredicate());
8881 if (Pred != CmpInst::BAD_ICMP_PREDICATE) {
8882 // If we are just checking for a icmp eq of a single bit and zext'ing it
8883 // to an integer, then shift the bit to the appropriate place and then
8884 // cast to integer to avoid the comparison.
8885 const APInt &Op1CV = CI->getValue();
8887 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
8888 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
8889 if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8890 (Pred == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
8891 Value *In = ICI->getOperand(0);
8892 Value *Sh = ConstantInt::get(In->getType(),
8893 In->getType()->getPrimitiveSizeInBits()-1);
8894 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
8895 In->getName()+".lobit"),
8897 if (In->getType() != SI.getType())
8898 In = CastInst::CreateIntegerCast(In, SI.getType(),
8899 true/*SExt*/, "tmp", ICI);
8901 if (Pred == ICmpInst::ICMP_SGT)
8902 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
8903 In->getName()+".not"), *ICI);
8905 return ReplaceInstUsesWith(SI, In);
8910 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
8911 // Transform (X == Y) ? X : Y -> Y
8912 if (Pred == ICmpInst::ICMP_EQ)
8913 return ReplaceInstUsesWith(SI, FalseVal);
8914 // Transform (X != Y) ? X : Y -> X
8915 if (Pred == ICmpInst::ICMP_NE)
8916 return ReplaceInstUsesWith(SI, TrueVal);
8917 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
8919 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
8920 // Transform (X == Y) ? Y : X -> X
8921 if (Pred == ICmpInst::ICMP_EQ)
8922 return ReplaceInstUsesWith(SI, FalseVal);
8923 // Transform (X != Y) ? Y : X -> Y
8924 if (Pred == ICmpInst::ICMP_NE)
8925 return ReplaceInstUsesWith(SI, TrueVal);
8926 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
8929 /// NOTE: if we wanted to, this is where to detect integer ABS
8931 return Changed ? &SI : 0;
8934 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
8935 Value *CondVal = SI.getCondition();
8936 Value *TrueVal = SI.getTrueValue();
8937 Value *FalseVal = SI.getFalseValue();
8939 // select true, X, Y -> X
8940 // select false, X, Y -> Y
8941 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
8942 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
8944 // select C, X, X -> X
8945 if (TrueVal == FalseVal)
8946 return ReplaceInstUsesWith(SI, TrueVal);
8948 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
8949 return ReplaceInstUsesWith(SI, FalseVal);
8950 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
8951 return ReplaceInstUsesWith(SI, TrueVal);
8952 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
8953 if (isa<Constant>(TrueVal))
8954 return ReplaceInstUsesWith(SI, TrueVal);
8956 return ReplaceInstUsesWith(SI, FalseVal);
8959 if (SI.getType() == Type::Int1Ty) {
8960 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
8961 if (C->getZExtValue()) {
8962 // Change: A = select B, true, C --> A = or B, C
8963 return BinaryOperator::CreateOr(CondVal, FalseVal);
8965 // Change: A = select B, false, C --> A = and !B, C
8967 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8968 "not."+CondVal->getName()), SI);
8969 return BinaryOperator::CreateAnd(NotCond, FalseVal);
8971 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
8972 if (C->getZExtValue() == false) {
8973 // Change: A = select B, C, false --> A = and B, C
8974 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8976 // Change: A = select B, C, true --> A = or !B, C
8978 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
8979 "not."+CondVal->getName()), SI);
8980 return BinaryOperator::CreateOr(NotCond, TrueVal);
8984 // select a, b, a -> a&b
8985 // select a, a, b -> a|b
8986 if (CondVal == TrueVal)
8987 return BinaryOperator::CreateOr(CondVal, FalseVal);
8988 else if (CondVal == FalseVal)
8989 return BinaryOperator::CreateAnd(CondVal, TrueVal);
8992 // Selecting between two integer constants?
8993 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
8994 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
8995 // select C, 1, 0 -> zext C to int
8996 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
8997 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
8998 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
8999 // select C, 0, 1 -> zext !C to int
9001 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9002 "not."+CondVal->getName()), SI);
9003 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
9006 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
9008 // (x <s 0) ? -1 : 0 -> ashr x, 31
9009 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
9010 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
9011 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
9012 // The comparison constant and the result are not neccessarily the
9013 // same width. Make an all-ones value by inserting a AShr.
9014 Value *X = IC->getOperand(0);
9015 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
9016 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
9017 Instruction *SRA = BinaryOperator::Create(Instruction::AShr, X,
9019 InsertNewInstBefore(SRA, SI);
9021 // Then cast to the appropriate width.
9022 return CastInst::CreateIntegerCast(SRA, SI.getType(), true);
9027 // If one of the constants is zero (we know they can't both be) and we
9028 // have an icmp instruction with zero, and we have an 'and' with the
9029 // non-constant value, eliminate this whole mess. This corresponds to
9030 // cases like this: ((X & 27) ? 27 : 0)
9031 if (TrueValC->isZero() || FalseValC->isZero())
9032 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
9033 cast<Constant>(IC->getOperand(1))->isNullValue())
9034 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
9035 if (ICA->getOpcode() == Instruction::And &&
9036 isa<ConstantInt>(ICA->getOperand(1)) &&
9037 (ICA->getOperand(1) == TrueValC ||
9038 ICA->getOperand(1) == FalseValC) &&
9039 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
9040 // Okay, now we know that everything is set up, we just don't
9041 // know whether we have a icmp_ne or icmp_eq and whether the
9042 // true or false val is the zero.
9043 bool ShouldNotVal = !TrueValC->isZero();
9044 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
9047 V = InsertNewInstBefore(BinaryOperator::Create(
9048 Instruction::Xor, V, ICA->getOperand(1)), SI);
9049 return ReplaceInstUsesWith(SI, V);
9054 // See if we are selecting two values based on a comparison of the two values.
9055 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
9056 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
9057 // Transform (X == Y) ? X : Y -> Y
9058 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9059 // This is not safe in general for floating point:
9060 // consider X== -0, Y== +0.
9061 // It becomes safe if either operand is a nonzero constant.
9062 ConstantFP *CFPt, *CFPf;
9063 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9064 !CFPt->getValueAPF().isZero()) ||
9065 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9066 !CFPf->getValueAPF().isZero()))
9067 return ReplaceInstUsesWith(SI, FalseVal);
9069 // Transform (X != Y) ? X : Y -> X
9070 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9071 return ReplaceInstUsesWith(SI, TrueVal);
9072 // NOTE: if we wanted to, this is where to detect MIN/MAX
9074 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
9075 // Transform (X == Y) ? Y : X -> X
9076 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9077 // This is not safe in general for floating point:
9078 // consider X== -0, Y== +0.
9079 // It becomes safe if either operand is a nonzero constant.
9080 ConstantFP *CFPt, *CFPf;
9081 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9082 !CFPt->getValueAPF().isZero()) ||
9083 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9084 !CFPf->getValueAPF().isZero()))
9085 return ReplaceInstUsesWith(SI, FalseVal);
9087 // Transform (X != Y) ? Y : X -> Y
9088 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9089 return ReplaceInstUsesWith(SI, TrueVal);
9090 // NOTE: if we wanted to, this is where to detect MIN/MAX
9092 // NOTE: if we wanted to, this is where to detect ABS
9095 // See if we are selecting two values based on a comparison of the two values.
9096 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
9097 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
9100 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
9101 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
9102 if (TI->hasOneUse() && FI->hasOneUse()) {
9103 Instruction *AddOp = 0, *SubOp = 0;
9105 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
9106 if (TI->getOpcode() == FI->getOpcode())
9107 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
9110 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
9111 // even legal for FP.
9112 if (TI->getOpcode() == Instruction::Sub &&
9113 FI->getOpcode() == Instruction::Add) {
9114 AddOp = FI; SubOp = TI;
9115 } else if (FI->getOpcode() == Instruction::Sub &&
9116 TI->getOpcode() == Instruction::Add) {
9117 AddOp = TI; SubOp = FI;
9121 Value *OtherAddOp = 0;
9122 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
9123 OtherAddOp = AddOp->getOperand(1);
9124 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
9125 OtherAddOp = AddOp->getOperand(0);
9129 // So at this point we know we have (Y -> OtherAddOp):
9130 // select C, (add X, Y), (sub X, Z)
9131 Value *NegVal; // Compute -Z
9132 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
9133 NegVal = ConstantExpr::getNeg(C);
9135 NegVal = InsertNewInstBefore(
9136 BinaryOperator::CreateNeg(SubOp->getOperand(1), "tmp"), SI);
9139 Value *NewTrueOp = OtherAddOp;
9140 Value *NewFalseOp = NegVal;
9142 std::swap(NewTrueOp, NewFalseOp);
9143 Instruction *NewSel =
9144 SelectInst::Create(CondVal, NewTrueOp,
9145 NewFalseOp, SI.getName() + ".p");
9147 NewSel = InsertNewInstBefore(NewSel, SI);
9148 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
9153 // See if we can fold the select into one of our operands.
9154 if (SI.getType()->isInteger()) {
9155 // See the comment above GetSelectFoldableOperands for a description of the
9156 // transformation we are doing here.
9157 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
9158 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
9159 !isa<Constant>(FalseVal))
9160 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
9161 unsigned OpToFold = 0;
9162 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
9164 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
9169 Constant *C = GetSelectFoldableConstant(TVI);
9170 Instruction *NewSel =
9171 SelectInst::Create(SI.getCondition(),
9172 TVI->getOperand(2-OpToFold), C);
9173 InsertNewInstBefore(NewSel, SI);
9174 NewSel->takeName(TVI);
9175 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
9176 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
9178 assert(0 && "Unknown instruction!!");
9183 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
9184 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
9185 !isa<Constant>(TrueVal))
9186 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
9187 unsigned OpToFold = 0;
9188 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
9190 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
9195 Constant *C = GetSelectFoldableConstant(FVI);
9196 Instruction *NewSel =
9197 SelectInst::Create(SI.getCondition(), C,
9198 FVI->getOperand(2-OpToFold));
9199 InsertNewInstBefore(NewSel, SI);
9200 NewSel->takeName(FVI);
9201 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
9202 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
9204 assert(0 && "Unknown instruction!!");
9209 if (BinaryOperator::isNot(CondVal)) {
9210 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
9211 SI.setOperand(1, FalseVal);
9212 SI.setOperand(2, TrueVal);
9219 /// EnforceKnownAlignment - If the specified pointer points to an object that
9220 /// we control, modify the object's alignment to PrefAlign. This isn't
9221 /// often possible though. If alignment is important, a more reliable approach
9222 /// is to simply align all global variables and allocation instructions to
9223 /// their preferred alignment from the beginning.
9225 static unsigned EnforceKnownAlignment(Value *V,
9226 unsigned Align, unsigned PrefAlign) {
9228 User *U = dyn_cast<User>(V);
9229 if (!U) return Align;
9231 switch (getOpcode(U)) {
9233 case Instruction::BitCast:
9234 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9235 case Instruction::GetElementPtr: {
9236 // If all indexes are zero, it is just the alignment of the base pointer.
9237 bool AllZeroOperands = true;
9238 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
9239 if (!isa<Constant>(*i) ||
9240 !cast<Constant>(*i)->isNullValue()) {
9241 AllZeroOperands = false;
9245 if (AllZeroOperands) {
9246 // Treat this like a bitcast.
9247 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9253 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
9254 // If there is a large requested alignment and we can, bump up the alignment
9256 if (!GV->isDeclaration()) {
9257 if (GV->getAlignment() >= PrefAlign)
9258 Align = GV->getAlignment();
9260 GV->setAlignment(PrefAlign);
9264 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
9265 // If there is a requested alignment and if this is an alloca, round up. We
9266 // don't do this for malloc, because some systems can't respect the request.
9267 if (isa<AllocaInst>(AI)) {
9268 if (AI->getAlignment() >= PrefAlign)
9269 Align = AI->getAlignment();
9271 AI->setAlignment(PrefAlign);
9280 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
9281 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
9282 /// and it is more than the alignment of the ultimate object, see if we can
9283 /// increase the alignment of the ultimate object, making this check succeed.
9284 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
9285 unsigned PrefAlign) {
9286 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
9287 sizeof(PrefAlign) * CHAR_BIT;
9288 APInt Mask = APInt::getAllOnesValue(BitWidth);
9289 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
9290 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
9291 unsigned TrailZ = KnownZero.countTrailingOnes();
9292 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
9294 if (PrefAlign > Align)
9295 Align = EnforceKnownAlignment(V, Align, PrefAlign);
9297 // We don't need to make any adjustment.
9301 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
9302 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
9303 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
9304 unsigned MinAlign = std::min(DstAlign, SrcAlign);
9305 unsigned CopyAlign = MI->getAlignment();
9307 if (CopyAlign < MinAlign) {
9308 MI->setAlignment(MinAlign);
9312 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
9314 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
9315 if (MemOpLength == 0) return 0;
9317 // Source and destination pointer types are always "i8*" for intrinsic. See
9318 // if the size is something we can handle with a single primitive load/store.
9319 // A single load+store correctly handles overlapping memory in the memmove
9321 unsigned Size = MemOpLength->getZExtValue();
9322 if (Size == 0) return MI; // Delete this mem transfer.
9324 if (Size > 8 || (Size&(Size-1)))
9325 return 0; // If not 1/2/4/8 bytes, exit.
9327 // Use an integer load+store unless we can find something better.
9328 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
9330 // Memcpy forces the use of i8* for the source and destination. That means
9331 // that if you're using memcpy to move one double around, you'll get a cast
9332 // from double* to i8*. We'd much rather use a double load+store rather than
9333 // an i64 load+store, here because this improves the odds that the source or
9334 // dest address will be promotable. See if we can find a better type than the
9335 // integer datatype.
9336 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
9337 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
9338 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
9339 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
9340 // down through these levels if so.
9341 while (!SrcETy->isSingleValueType()) {
9342 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
9343 if (STy->getNumElements() == 1)
9344 SrcETy = STy->getElementType(0);
9347 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
9348 if (ATy->getNumElements() == 1)
9349 SrcETy = ATy->getElementType();
9356 if (SrcETy->isSingleValueType())
9357 NewPtrTy = PointerType::getUnqual(SrcETy);
9362 // If the memcpy/memmove provides better alignment info than we can
9364 SrcAlign = std::max(SrcAlign, CopyAlign);
9365 DstAlign = std::max(DstAlign, CopyAlign);
9367 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
9368 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
9369 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
9370 InsertNewInstBefore(L, *MI);
9371 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
9373 // Set the size of the copy to 0, it will be deleted on the next iteration.
9374 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
9378 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
9379 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
9380 if (MI->getAlignment() < Alignment) {
9381 MI->setAlignment(Alignment);
9385 // Extract the length and alignment and fill if they are constant.
9386 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
9387 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
9388 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
9390 uint64_t Len = LenC->getZExtValue();
9391 Alignment = MI->getAlignment();
9393 // If the length is zero, this is a no-op
9394 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
9396 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
9397 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
9398 const Type *ITy = IntegerType::get(Len*8); // n=1 -> i8.
9400 Value *Dest = MI->getDest();
9401 Dest = InsertBitCastBefore(Dest, PointerType::getUnqual(ITy), *MI);
9403 // Alignment 0 is identity for alignment 1 for memset, but not store.
9404 if (Alignment == 0) Alignment = 1;
9406 // Extract the fill value and store.
9407 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
9408 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill), Dest, false,
9411 // Set the size of the copy to 0, it will be deleted on the next iteration.
9412 MI->setLength(Constant::getNullValue(LenC->getType()));
9420 /// visitCallInst - CallInst simplification. This mostly only handles folding
9421 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
9422 /// the heavy lifting.
9424 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
9425 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
9426 if (!II) return visitCallSite(&CI);
9428 // Intrinsics cannot occur in an invoke, so handle them here instead of in
9430 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
9431 bool Changed = false;
9433 // memmove/cpy/set of zero bytes is a noop.
9434 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
9435 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
9437 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
9438 if (CI->getZExtValue() == 1) {
9439 // Replace the instruction with just byte operations. We would
9440 // transform other cases to loads/stores, but we don't know if
9441 // alignment is sufficient.
9445 // If we have a memmove and the source operation is a constant global,
9446 // then the source and dest pointers can't alias, so we can change this
9447 // into a call to memcpy.
9448 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
9449 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
9450 if (GVSrc->isConstant()) {
9451 Module *M = CI.getParent()->getParent()->getParent();
9452 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
9454 Tys[0] = CI.getOperand(3)->getType();
9456 Intrinsic::getDeclaration(M, MemCpyID, Tys, 1));
9460 // memmove(x,x,size) -> noop.
9461 if (MMI->getSource() == MMI->getDest())
9462 return EraseInstFromFunction(CI);
9465 // If we can determine a pointer alignment that is bigger than currently
9466 // set, update the alignment.
9467 if (isa<MemTransferInst>(MI)) {
9468 if (Instruction *I = SimplifyMemTransfer(MI))
9470 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
9471 if (Instruction *I = SimplifyMemSet(MSI))
9475 if (Changed) return II;
9478 switch (II->getIntrinsicID()) {
9480 case Intrinsic::bswap:
9481 // bswap(bswap(x)) -> x
9482 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
9483 if (Operand->getIntrinsicID() == Intrinsic::bswap)
9484 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
9486 case Intrinsic::ppc_altivec_lvx:
9487 case Intrinsic::ppc_altivec_lvxl:
9488 case Intrinsic::x86_sse_loadu_ps:
9489 case Intrinsic::x86_sse2_loadu_pd:
9490 case Intrinsic::x86_sse2_loadu_dq:
9491 // Turn PPC lvx -> load if the pointer is known aligned.
9492 // Turn X86 loadups -> load if the pointer is known aligned.
9493 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9494 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
9495 PointerType::getUnqual(II->getType()),
9497 return new LoadInst(Ptr);
9500 case Intrinsic::ppc_altivec_stvx:
9501 case Intrinsic::ppc_altivec_stvxl:
9502 // Turn stvx -> store if the pointer is known aligned.
9503 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
9504 const Type *OpPtrTy =
9505 PointerType::getUnqual(II->getOperand(1)->getType());
9506 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
9507 return new StoreInst(II->getOperand(1), Ptr);
9510 case Intrinsic::x86_sse_storeu_ps:
9511 case Intrinsic::x86_sse2_storeu_pd:
9512 case Intrinsic::x86_sse2_storeu_dq:
9513 // Turn X86 storeu -> store if the pointer is known aligned.
9514 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9515 const Type *OpPtrTy =
9516 PointerType::getUnqual(II->getOperand(2)->getType());
9517 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
9518 return new StoreInst(II->getOperand(2), Ptr);
9522 case Intrinsic::x86_sse_cvttss2si: {
9523 // These intrinsics only demands the 0th element of its input vector. If
9524 // we can simplify the input based on that, do so now.
9526 cast<VectorType>(II->getOperand(1)->getType())->getNumElements();
9527 APInt DemandedElts(VWidth, 1);
9528 APInt UndefElts(VWidth, 0);
9529 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
9531 II->setOperand(1, V);
9537 case Intrinsic::ppc_altivec_vperm:
9538 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
9539 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
9540 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
9542 // Check that all of the elements are integer constants or undefs.
9543 bool AllEltsOk = true;
9544 for (unsigned i = 0; i != 16; ++i) {
9545 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
9546 !isa<UndefValue>(Mask->getOperand(i))) {
9553 // Cast the input vectors to byte vectors.
9554 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
9555 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
9556 Value *Result = UndefValue::get(Op0->getType());
9558 // Only extract each element once.
9559 Value *ExtractedElts[32];
9560 memset(ExtractedElts, 0, sizeof(ExtractedElts));
9562 for (unsigned i = 0; i != 16; ++i) {
9563 if (isa<UndefValue>(Mask->getOperand(i)))
9565 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
9566 Idx &= 31; // Match the hardware behavior.
9568 if (ExtractedElts[Idx] == 0) {
9570 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
9571 InsertNewInstBefore(Elt, CI);
9572 ExtractedElts[Idx] = Elt;
9575 // Insert this value into the result vector.
9576 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
9578 InsertNewInstBefore(cast<Instruction>(Result), CI);
9580 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
9585 case Intrinsic::stackrestore: {
9586 // If the save is right next to the restore, remove the restore. This can
9587 // happen when variable allocas are DCE'd.
9588 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
9589 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
9590 BasicBlock::iterator BI = SS;
9592 return EraseInstFromFunction(CI);
9596 // Scan down this block to see if there is another stack restore in the
9597 // same block without an intervening call/alloca.
9598 BasicBlock::iterator BI = II;
9599 TerminatorInst *TI = II->getParent()->getTerminator();
9600 bool CannotRemove = false;
9601 for (++BI; &*BI != TI; ++BI) {
9602 if (isa<AllocaInst>(BI)) {
9603 CannotRemove = true;
9606 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
9607 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
9608 // If there is a stackrestore below this one, remove this one.
9609 if (II->getIntrinsicID() == Intrinsic::stackrestore)
9610 return EraseInstFromFunction(CI);
9611 // Otherwise, ignore the intrinsic.
9613 // If we found a non-intrinsic call, we can't remove the stack
9615 CannotRemove = true;
9621 // If the stack restore is in a return/unwind block and if there are no
9622 // allocas or calls between the restore and the return, nuke the restore.
9623 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
9624 return EraseInstFromFunction(CI);
9629 return visitCallSite(II);
9632 // InvokeInst simplification
9634 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
9635 return visitCallSite(&II);
9638 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
9639 /// passed through the varargs area, we can eliminate the use of the cast.
9640 static bool isSafeToEliminateVarargsCast(const CallSite CS,
9641 const CastInst * const CI,
9642 const TargetData * const TD,
9644 if (!CI->isLosslessCast())
9647 // The size of ByVal arguments is derived from the type, so we
9648 // can't change to a type with a different size. If the size were
9649 // passed explicitly we could avoid this check.
9650 if (!CS.paramHasAttr(ix, Attribute::ByVal))
9654 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
9655 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
9656 if (!SrcTy->isSized() || !DstTy->isSized())
9658 if (TD->getTypePaddedSize(SrcTy) != TD->getTypePaddedSize(DstTy))
9663 // visitCallSite - Improvements for call and invoke instructions.
9665 Instruction *InstCombiner::visitCallSite(CallSite CS) {
9666 bool Changed = false;
9668 // If the callee is a constexpr cast of a function, attempt to move the cast
9669 // to the arguments of the call/invoke.
9670 if (transformConstExprCastCall(CS)) return 0;
9672 Value *Callee = CS.getCalledValue();
9674 if (Function *CalleeF = dyn_cast<Function>(Callee))
9675 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
9676 Instruction *OldCall = CS.getInstruction();
9677 // If the call and callee calling conventions don't match, this call must
9678 // be unreachable, as the call is undefined.
9679 new StoreInst(ConstantInt::getTrue(),
9680 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9682 if (!OldCall->use_empty())
9683 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
9684 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
9685 return EraseInstFromFunction(*OldCall);
9689 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
9690 // This instruction is not reachable, just remove it. We insert a store to
9691 // undef so that we know that this code is not reachable, despite the fact
9692 // that we can't modify the CFG here.
9693 new StoreInst(ConstantInt::getTrue(),
9694 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
9695 CS.getInstruction());
9697 if (!CS.getInstruction()->use_empty())
9698 CS.getInstruction()->
9699 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
9701 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
9702 // Don't break the CFG, insert a dummy cond branch.
9703 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
9704 ConstantInt::getTrue(), II);
9706 return EraseInstFromFunction(*CS.getInstruction());
9709 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
9710 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
9711 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
9712 return transformCallThroughTrampoline(CS);
9714 const PointerType *PTy = cast<PointerType>(Callee->getType());
9715 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9716 if (FTy->isVarArg()) {
9717 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
9718 // See if we can optimize any arguments passed through the varargs area of
9720 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
9721 E = CS.arg_end(); I != E; ++I, ++ix) {
9722 CastInst *CI = dyn_cast<CastInst>(*I);
9723 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
9724 *I = CI->getOperand(0);
9730 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
9731 // Inline asm calls cannot throw - mark them 'nounwind'.
9732 CS.setDoesNotThrow();
9736 return Changed ? CS.getInstruction() : 0;
9739 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
9740 // attempt to move the cast to the arguments of the call/invoke.
9742 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
9743 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
9744 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
9745 if (CE->getOpcode() != Instruction::BitCast ||
9746 !isa<Function>(CE->getOperand(0)))
9748 Function *Callee = cast<Function>(CE->getOperand(0));
9749 Instruction *Caller = CS.getInstruction();
9750 const AttrListPtr &CallerPAL = CS.getAttributes();
9752 // Okay, this is a cast from a function to a different type. Unless doing so
9753 // would cause a type conversion of one of our arguments, change this call to
9754 // be a direct call with arguments casted to the appropriate types.
9756 const FunctionType *FT = Callee->getFunctionType();
9757 const Type *OldRetTy = Caller->getType();
9758 const Type *NewRetTy = FT->getReturnType();
9760 if (isa<StructType>(NewRetTy))
9761 return false; // TODO: Handle multiple return values.
9763 // Check to see if we are changing the return type...
9764 if (OldRetTy != NewRetTy) {
9765 if (Callee->isDeclaration() &&
9766 // Conversion is ok if changing from one pointer type to another or from
9767 // a pointer to an integer of the same size.
9768 !((isa<PointerType>(OldRetTy) || OldRetTy == TD->getIntPtrType()) &&
9769 (isa<PointerType>(NewRetTy) || NewRetTy == TD->getIntPtrType())))
9770 return false; // Cannot transform this return value.
9772 if (!Caller->use_empty() &&
9773 // void -> non-void is handled specially
9774 NewRetTy != Type::VoidTy && !CastInst::isCastable(NewRetTy, OldRetTy))
9775 return false; // Cannot transform this return value.
9777 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
9778 Attributes RAttrs = CallerPAL.getRetAttributes();
9779 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
9780 return false; // Attribute not compatible with transformed value.
9783 // If the callsite is an invoke instruction, and the return value is used by
9784 // a PHI node in a successor, we cannot change the return type of the call
9785 // because there is no place to put the cast instruction (without breaking
9786 // the critical edge). Bail out in this case.
9787 if (!Caller->use_empty())
9788 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
9789 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
9791 if (PHINode *PN = dyn_cast<PHINode>(*UI))
9792 if (PN->getParent() == II->getNormalDest() ||
9793 PN->getParent() == II->getUnwindDest())
9797 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
9798 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
9800 CallSite::arg_iterator AI = CS.arg_begin();
9801 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
9802 const Type *ParamTy = FT->getParamType(i);
9803 const Type *ActTy = (*AI)->getType();
9805 if (!CastInst::isCastable(ActTy, ParamTy))
9806 return false; // Cannot transform this parameter value.
9808 if (CallerPAL.getParamAttributes(i + 1)
9809 & Attribute::typeIncompatible(ParamTy))
9810 return false; // Attribute not compatible with transformed value.
9812 // Converting from one pointer type to another or between a pointer and an
9813 // integer of the same size is safe even if we do not have a body.
9814 bool isConvertible = ActTy == ParamTy ||
9815 ((isa<PointerType>(ParamTy) || ParamTy == TD->getIntPtrType()) &&
9816 (isa<PointerType>(ActTy) || ActTy == TD->getIntPtrType()));
9817 if (Callee->isDeclaration() && !isConvertible) return false;
9820 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
9821 Callee->isDeclaration())
9822 return false; // Do not delete arguments unless we have a function body.
9824 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
9825 !CallerPAL.isEmpty())
9826 // In this case we have more arguments than the new function type, but we
9827 // won't be dropping them. Check that these extra arguments have attributes
9828 // that are compatible with being a vararg call argument.
9829 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
9830 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
9832 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
9833 if (PAttrs & Attribute::VarArgsIncompatible)
9837 // Okay, we decided that this is a safe thing to do: go ahead and start
9838 // inserting cast instructions as necessary...
9839 std::vector<Value*> Args;
9840 Args.reserve(NumActualArgs);
9841 SmallVector<AttributeWithIndex, 8> attrVec;
9842 attrVec.reserve(NumCommonArgs);
9844 // Get any return attributes.
9845 Attributes RAttrs = CallerPAL.getRetAttributes();
9847 // If the return value is not being used, the type may not be compatible
9848 // with the existing attributes. Wipe out any problematic attributes.
9849 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
9851 // Add the new return attributes.
9853 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
9855 AI = CS.arg_begin();
9856 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
9857 const Type *ParamTy = FT->getParamType(i);
9858 if ((*AI)->getType() == ParamTy) {
9859 Args.push_back(*AI);
9861 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
9862 false, ParamTy, false);
9863 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
9864 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
9867 // Add any parameter attributes.
9868 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
9869 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
9872 // If the function takes more arguments than the call was taking, add them
9874 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
9875 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
9877 // If we are removing arguments to the function, emit an obnoxious warning...
9878 if (FT->getNumParams() < NumActualArgs) {
9879 if (!FT->isVarArg()) {
9880 cerr << "WARNING: While resolving call to function '"
9881 << Callee->getName() << "' arguments were dropped!\n";
9883 // Add all of the arguments in their promoted form to the arg list...
9884 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
9885 const Type *PTy = getPromotedType((*AI)->getType());
9886 if (PTy != (*AI)->getType()) {
9887 // Must promote to pass through va_arg area!
9888 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
9890 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
9891 InsertNewInstBefore(Cast, *Caller);
9892 Args.push_back(Cast);
9894 Args.push_back(*AI);
9897 // Add any parameter attributes.
9898 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
9899 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
9904 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
9905 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
9907 if (NewRetTy == Type::VoidTy)
9908 Caller->setName(""); // Void type should not have a name.
9910 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),attrVec.end());
9913 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9914 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
9915 Args.begin(), Args.end(),
9916 Caller->getName(), Caller);
9917 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
9918 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
9920 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
9921 Caller->getName(), Caller);
9922 CallInst *CI = cast<CallInst>(Caller);
9923 if (CI->isTailCall())
9924 cast<CallInst>(NC)->setTailCall();
9925 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
9926 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
9929 // Insert a cast of the return type as necessary.
9931 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
9932 if (NV->getType() != Type::VoidTy) {
9933 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
9935 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
9937 // If this is an invoke instruction, we should insert it after the first
9938 // non-phi, instruction in the normal successor block.
9939 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9940 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
9941 InsertNewInstBefore(NC, *I);
9943 // Otherwise, it's a call, just insert cast right after the call instr
9944 InsertNewInstBefore(NC, *Caller);
9946 AddUsersToWorkList(*Caller);
9948 NV = UndefValue::get(Caller->getType());
9952 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9953 Caller->replaceAllUsesWith(NV);
9954 Caller->eraseFromParent();
9955 RemoveFromWorkList(Caller);
9959 // transformCallThroughTrampoline - Turn a call to a function created by the
9960 // init_trampoline intrinsic into a direct call to the underlying function.
9962 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
9963 Value *Callee = CS.getCalledValue();
9964 const PointerType *PTy = cast<PointerType>(Callee->getType());
9965 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9966 const AttrListPtr &Attrs = CS.getAttributes();
9968 // If the call already has the 'nest' attribute somewhere then give up -
9969 // otherwise 'nest' would occur twice after splicing in the chain.
9970 if (Attrs.hasAttrSomewhere(Attribute::Nest))
9973 IntrinsicInst *Tramp =
9974 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
9976 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
9977 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
9978 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
9980 const AttrListPtr &NestAttrs = NestF->getAttributes();
9981 if (!NestAttrs.isEmpty()) {
9982 unsigned NestIdx = 1;
9983 const Type *NestTy = 0;
9984 Attributes NestAttr = Attribute::None;
9986 // Look for a parameter marked with the 'nest' attribute.
9987 for (FunctionType::param_iterator I = NestFTy->param_begin(),
9988 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
9989 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
9990 // Record the parameter type and any other attributes.
9992 NestAttr = NestAttrs.getParamAttributes(NestIdx);
9997 Instruction *Caller = CS.getInstruction();
9998 std::vector<Value*> NewArgs;
9999 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
10001 SmallVector<AttributeWithIndex, 8> NewAttrs;
10002 NewAttrs.reserve(Attrs.getNumSlots() + 1);
10004 // Insert the nest argument into the call argument list, which may
10005 // mean appending it. Likewise for attributes.
10007 // Add any result attributes.
10008 if (Attributes Attr = Attrs.getRetAttributes())
10009 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
10013 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
10015 if (Idx == NestIdx) {
10016 // Add the chain argument and attributes.
10017 Value *NestVal = Tramp->getOperand(3);
10018 if (NestVal->getType() != NestTy)
10019 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
10020 NewArgs.push_back(NestVal);
10021 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
10027 // Add the original argument and attributes.
10028 NewArgs.push_back(*I);
10029 if (Attributes Attr = Attrs.getParamAttributes(Idx))
10031 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
10037 // Add any function attributes.
10038 if (Attributes Attr = Attrs.getFnAttributes())
10039 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
10041 // The trampoline may have been bitcast to a bogus type (FTy).
10042 // Handle this by synthesizing a new function type, equal to FTy
10043 // with the chain parameter inserted.
10045 std::vector<const Type*> NewTypes;
10046 NewTypes.reserve(FTy->getNumParams()+1);
10048 // Insert the chain's type into the list of parameter types, which may
10049 // mean appending it.
10052 FunctionType::param_iterator I = FTy->param_begin(),
10053 E = FTy->param_end();
10056 if (Idx == NestIdx)
10057 // Add the chain's type.
10058 NewTypes.push_back(NestTy);
10063 // Add the original type.
10064 NewTypes.push_back(*I);
10070 // Replace the trampoline call with a direct call. Let the generic
10071 // code sort out any function type mismatches.
10072 FunctionType *NewFTy =
10073 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
10074 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
10075 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
10076 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),NewAttrs.end());
10078 Instruction *NewCaller;
10079 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10080 NewCaller = InvokeInst::Create(NewCallee,
10081 II->getNormalDest(), II->getUnwindDest(),
10082 NewArgs.begin(), NewArgs.end(),
10083 Caller->getName(), Caller);
10084 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
10085 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
10087 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
10088 Caller->getName(), Caller);
10089 if (cast<CallInst>(Caller)->isTailCall())
10090 cast<CallInst>(NewCaller)->setTailCall();
10091 cast<CallInst>(NewCaller)->
10092 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
10093 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
10095 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
10096 Caller->replaceAllUsesWith(NewCaller);
10097 Caller->eraseFromParent();
10098 RemoveFromWorkList(Caller);
10103 // Replace the trampoline call with a direct call. Since there is no 'nest'
10104 // parameter, there is no need to adjust the argument list. Let the generic
10105 // code sort out any function type mismatches.
10106 Constant *NewCallee =
10107 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
10108 CS.setCalledFunction(NewCallee);
10109 return CS.getInstruction();
10112 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
10113 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
10114 /// and a single binop.
10115 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
10116 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10117 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
10118 unsigned Opc = FirstInst->getOpcode();
10119 Value *LHSVal = FirstInst->getOperand(0);
10120 Value *RHSVal = FirstInst->getOperand(1);
10122 const Type *LHSType = LHSVal->getType();
10123 const Type *RHSType = RHSVal->getType();
10125 // Scan to see if all operands are the same opcode, all have one use, and all
10126 // kill their operands (i.e. the operands have one use).
10127 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10128 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
10129 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
10130 // Verify type of the LHS matches so we don't fold cmp's of different
10131 // types or GEP's with different index types.
10132 I->getOperand(0)->getType() != LHSType ||
10133 I->getOperand(1)->getType() != RHSType)
10136 // If they are CmpInst instructions, check their predicates
10137 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
10138 if (cast<CmpInst>(I)->getPredicate() !=
10139 cast<CmpInst>(FirstInst)->getPredicate())
10142 // Keep track of which operand needs a phi node.
10143 if (I->getOperand(0) != LHSVal) LHSVal = 0;
10144 if (I->getOperand(1) != RHSVal) RHSVal = 0;
10147 // Otherwise, this is safe to transform!
10149 Value *InLHS = FirstInst->getOperand(0);
10150 Value *InRHS = FirstInst->getOperand(1);
10151 PHINode *NewLHS = 0, *NewRHS = 0;
10153 NewLHS = PHINode::Create(LHSType,
10154 FirstInst->getOperand(0)->getName() + ".pn");
10155 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
10156 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
10157 InsertNewInstBefore(NewLHS, PN);
10162 NewRHS = PHINode::Create(RHSType,
10163 FirstInst->getOperand(1)->getName() + ".pn");
10164 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
10165 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
10166 InsertNewInstBefore(NewRHS, PN);
10170 // Add all operands to the new PHIs.
10171 if (NewLHS || NewRHS) {
10172 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10173 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
10175 Value *NewInLHS = InInst->getOperand(0);
10176 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
10179 Value *NewInRHS = InInst->getOperand(1);
10180 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
10185 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10186 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
10187 CmpInst *CIOp = cast<CmpInst>(FirstInst);
10188 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
10192 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) {
10193 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
10195 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
10196 FirstInst->op_end());
10197 // This is true if all GEP bases are allocas and if all indices into them are
10199 bool AllBasePointersAreAllocas = true;
10201 // Scan to see if all operands are the same opcode, all have one use, and all
10202 // kill their operands (i.e. the operands have one use).
10203 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10204 GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
10205 if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() ||
10206 GEP->getNumOperands() != FirstInst->getNumOperands())
10209 // Keep track of whether or not all GEPs are of alloca pointers.
10210 if (AllBasePointersAreAllocas &&
10211 (!isa<AllocaInst>(GEP->getOperand(0)) ||
10212 !GEP->hasAllConstantIndices()))
10213 AllBasePointersAreAllocas = false;
10215 // Compare the operand lists.
10216 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
10217 if (FirstInst->getOperand(op) == GEP->getOperand(op))
10220 // Don't merge two GEPs when two operands differ (introducing phi nodes)
10221 // if one of the PHIs has a constant for the index. The index may be
10222 // substantially cheaper to compute for the constants, so making it a
10223 // variable index could pessimize the path. This also handles the case
10224 // for struct indices, which must always be constant.
10225 if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
10226 isa<ConstantInt>(GEP->getOperand(op)))
10229 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
10231 FixedOperands[op] = 0; // Needs a PHI.
10235 // If all of the base pointers of the PHI'd GEPs are from allocas, don't
10236 // bother doing this transformation. At best, this will just save a bit of
10237 // offset calculation, but all the predecessors will have to materialize the
10238 // stack address into a register anyway. We'd actually rather *clone* the
10239 // load up into the predecessors so that we have a load of a gep of an alloca,
10240 // which can usually all be folded into the load.
10241 if (AllBasePointersAreAllocas)
10244 // Otherwise, this is safe to transform. Insert PHI nodes for each operand
10245 // that is variable.
10246 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
10248 bool HasAnyPHIs = false;
10249 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
10250 if (FixedOperands[i]) continue; // operand doesn't need a phi.
10251 Value *FirstOp = FirstInst->getOperand(i);
10252 PHINode *NewPN = PHINode::Create(FirstOp->getType(),
10253 FirstOp->getName()+".pn");
10254 InsertNewInstBefore(NewPN, PN);
10256 NewPN->reserveOperandSpace(e);
10257 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
10258 OperandPhis[i] = NewPN;
10259 FixedOperands[i] = NewPN;
10264 // Add all operands to the new PHIs.
10266 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10267 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
10268 BasicBlock *InBB = PN.getIncomingBlock(i);
10270 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
10271 if (PHINode *OpPhi = OperandPhis[op])
10272 OpPhi->addIncoming(InGEP->getOperand(op), InBB);
10276 Value *Base = FixedOperands[0];
10277 return GetElementPtrInst::Create(Base, FixedOperands.begin()+1,
10278 FixedOperands.end());
10282 /// isSafeAndProfitableToSinkLoad - Return true if we know that it is safe to
10283 /// sink the load out of the block that defines it. This means that it must be
10284 /// obvious the value of the load is not changed from the point of the load to
10285 /// the end of the block it is in.
10287 /// Finally, it is safe, but not profitable, to sink a load targetting a
10288 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
10290 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {
10291 BasicBlock::iterator BBI = L, E = L->getParent()->end();
10293 for (++BBI; BBI != E; ++BBI)
10294 if (BBI->mayWriteToMemory())
10297 // Check for non-address taken alloca. If not address-taken already, it isn't
10298 // profitable to do this xform.
10299 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
10300 bool isAddressTaken = false;
10301 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
10303 if (isa<LoadInst>(UI)) continue;
10304 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
10305 // If storing TO the alloca, then the address isn't taken.
10306 if (SI->getOperand(1) == AI) continue;
10308 isAddressTaken = true;
10312 if (!isAddressTaken && AI->isStaticAlloca())
10316 // If this load is a load from a GEP with a constant offset from an alloca,
10317 // then we don't want to sink it. In its present form, it will be
10318 // load [constant stack offset]. Sinking it will cause us to have to
10319 // materialize the stack addresses in each predecessor in a register only to
10320 // do a shared load from register in the successor.
10321 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0)))
10322 if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0)))
10323 if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())
10330 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
10331 // operator and they all are only used by the PHI, PHI together their
10332 // inputs, and do the operation once, to the result of the PHI.
10333 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
10334 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10336 // Scan the instruction, looking for input operations that can be folded away.
10337 // If all input operands to the phi are the same instruction (e.g. a cast from
10338 // the same type or "+42") we can pull the operation through the PHI, reducing
10339 // code size and simplifying code.
10340 Constant *ConstantOp = 0;
10341 const Type *CastSrcTy = 0;
10342 bool isVolatile = false;
10343 if (isa<CastInst>(FirstInst)) {
10344 CastSrcTy = FirstInst->getOperand(0)->getType();
10345 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
10346 // Can fold binop, compare or shift here if the RHS is a constant,
10347 // otherwise call FoldPHIArgBinOpIntoPHI.
10348 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
10349 if (ConstantOp == 0)
10350 return FoldPHIArgBinOpIntoPHI(PN);
10351 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
10352 isVolatile = LI->isVolatile();
10353 // We can't sink the load if the loaded value could be modified between the
10354 // load and the PHI.
10355 if (LI->getParent() != PN.getIncomingBlock(0) ||
10356 !isSafeAndProfitableToSinkLoad(LI))
10359 // If the PHI is of volatile loads and the load block has multiple
10360 // successors, sinking it would remove a load of the volatile value from
10361 // the path through the other successor.
10363 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10366 } else if (isa<GetElementPtrInst>(FirstInst)) {
10367 return FoldPHIArgGEPIntoPHI(PN);
10369 return 0; // Cannot fold this operation.
10372 // Check to see if all arguments are the same operation.
10373 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10374 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
10375 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
10376 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
10379 if (I->getOperand(0)->getType() != CastSrcTy)
10380 return 0; // Cast operation must match.
10381 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10382 // We can't sink the load if the loaded value could be modified between
10383 // the load and the PHI.
10384 if (LI->isVolatile() != isVolatile ||
10385 LI->getParent() != PN.getIncomingBlock(i) ||
10386 !isSafeAndProfitableToSinkLoad(LI))
10389 // If the PHI is of volatile loads and the load block has multiple
10390 // successors, sinking it would remove a load of the volatile value from
10391 // the path through the other successor.
10393 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10396 } else if (I->getOperand(1) != ConstantOp) {
10401 // Okay, they are all the same operation. Create a new PHI node of the
10402 // correct type, and PHI together all of the LHS's of the instructions.
10403 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
10404 PN.getName()+".in");
10405 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
10407 Value *InVal = FirstInst->getOperand(0);
10408 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
10410 // Add all operands to the new PHI.
10411 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10412 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
10413 if (NewInVal != InVal)
10415 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
10420 // The new PHI unions all of the same values together. This is really
10421 // common, so we handle it intelligently here for compile-time speed.
10425 InsertNewInstBefore(NewPN, PN);
10429 // Insert and return the new operation.
10430 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
10431 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
10432 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10433 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
10434 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
10435 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
10436 PhiVal, ConstantOp);
10437 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
10439 // If this was a volatile load that we are merging, make sure to loop through
10440 // and mark all the input loads as non-volatile. If we don't do this, we will
10441 // insert a new volatile load and the old ones will not be deletable.
10443 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
10444 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
10446 return new LoadInst(PhiVal, "", isVolatile);
10449 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
10451 static bool DeadPHICycle(PHINode *PN,
10452 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
10453 if (PN->use_empty()) return true;
10454 if (!PN->hasOneUse()) return false;
10456 // Remember this node, and if we find the cycle, return.
10457 if (!PotentiallyDeadPHIs.insert(PN))
10460 // Don't scan crazily complex things.
10461 if (PotentiallyDeadPHIs.size() == 16)
10464 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
10465 return DeadPHICycle(PU, PotentiallyDeadPHIs);
10470 /// PHIsEqualValue - Return true if this phi node is always equal to
10471 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
10472 /// z = some value; x = phi (y, z); y = phi (x, z)
10473 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
10474 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
10475 // See if we already saw this PHI node.
10476 if (!ValueEqualPHIs.insert(PN))
10479 // Don't scan crazily complex things.
10480 if (ValueEqualPHIs.size() == 16)
10483 // Scan the operands to see if they are either phi nodes or are equal to
10485 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
10486 Value *Op = PN->getIncomingValue(i);
10487 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
10488 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
10490 } else if (Op != NonPhiInVal)
10498 // PHINode simplification
10500 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
10501 // If LCSSA is around, don't mess with Phi nodes
10502 if (MustPreserveLCSSA) return 0;
10504 if (Value *V = PN.hasConstantValue())
10505 return ReplaceInstUsesWith(PN, V);
10507 // If all PHI operands are the same operation, pull them through the PHI,
10508 // reducing code size.
10509 if (isa<Instruction>(PN.getIncomingValue(0)) &&
10510 isa<Instruction>(PN.getIncomingValue(1)) &&
10511 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
10512 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
10513 // FIXME: The hasOneUse check will fail for PHIs that use the value more
10514 // than themselves more than once.
10515 PN.getIncomingValue(0)->hasOneUse())
10516 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
10519 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
10520 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
10521 // PHI)... break the cycle.
10522 if (PN.hasOneUse()) {
10523 Instruction *PHIUser = cast<Instruction>(PN.use_back());
10524 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
10525 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
10526 PotentiallyDeadPHIs.insert(&PN);
10527 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
10528 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10531 // If this phi has a single use, and if that use just computes a value for
10532 // the next iteration of a loop, delete the phi. This occurs with unused
10533 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
10534 // common case here is good because the only other things that catch this
10535 // are induction variable analysis (sometimes) and ADCE, which is only run
10537 if (PHIUser->hasOneUse() &&
10538 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
10539 PHIUser->use_back() == &PN) {
10540 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10544 // We sometimes end up with phi cycles that non-obviously end up being the
10545 // same value, for example:
10546 // z = some value; x = phi (y, z); y = phi (x, z)
10547 // where the phi nodes don't necessarily need to be in the same block. Do a
10548 // quick check to see if the PHI node only contains a single non-phi value, if
10549 // so, scan to see if the phi cycle is actually equal to that value.
10551 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
10552 // Scan for the first non-phi operand.
10553 while (InValNo != NumOperandVals &&
10554 isa<PHINode>(PN.getIncomingValue(InValNo)))
10557 if (InValNo != NumOperandVals) {
10558 Value *NonPhiInVal = PN.getOperand(InValNo);
10560 // Scan the rest of the operands to see if there are any conflicts, if so
10561 // there is no need to recursively scan other phis.
10562 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
10563 Value *OpVal = PN.getIncomingValue(InValNo);
10564 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
10568 // If we scanned over all operands, then we have one unique value plus
10569 // phi values. Scan PHI nodes to see if they all merge in each other or
10571 if (InValNo == NumOperandVals) {
10572 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
10573 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
10574 return ReplaceInstUsesWith(PN, NonPhiInVal);
10581 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
10582 Instruction *InsertPoint,
10583 InstCombiner *IC) {
10584 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
10585 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
10586 // We must cast correctly to the pointer type. Ensure that we
10587 // sign extend the integer value if it is smaller as this is
10588 // used for address computation.
10589 Instruction::CastOps opcode =
10590 (VTySize < PtrSize ? Instruction::SExt :
10591 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
10592 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
10596 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
10597 Value *PtrOp = GEP.getOperand(0);
10598 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
10599 // If so, eliminate the noop.
10600 if (GEP.getNumOperands() == 1)
10601 return ReplaceInstUsesWith(GEP, PtrOp);
10603 if (isa<UndefValue>(GEP.getOperand(0)))
10604 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
10606 bool HasZeroPointerIndex = false;
10607 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
10608 HasZeroPointerIndex = C->isNullValue();
10610 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
10611 return ReplaceInstUsesWith(GEP, PtrOp);
10613 // Eliminate unneeded casts for indices.
10614 bool MadeChange = false;
10616 gep_type_iterator GTI = gep_type_begin(GEP);
10617 for (User::op_iterator i = GEP.op_begin() + 1, e = GEP.op_end();
10618 i != e; ++i, ++GTI) {
10619 if (isa<SequentialType>(*GTI)) {
10620 if (CastInst *CI = dyn_cast<CastInst>(*i)) {
10621 if (CI->getOpcode() == Instruction::ZExt ||
10622 CI->getOpcode() == Instruction::SExt) {
10623 const Type *SrcTy = CI->getOperand(0)->getType();
10624 // We can eliminate a cast from i32 to i64 iff the target
10625 // is a 32-bit pointer target.
10626 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
10628 *i = CI->getOperand(0);
10632 // If we are using a wider index than needed for this platform, shrink it
10633 // to what we need. If narrower, sign-extend it to what we need.
10634 // If the incoming value needs a cast instruction,
10635 // insert it. This explicit cast can make subsequent optimizations more
10638 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
10639 if (Constant *C = dyn_cast<Constant>(Op)) {
10640 *i = ConstantExpr::getTrunc(C, TD->getIntPtrType());
10643 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
10648 } else if (TD->getTypeSizeInBits(Op->getType()) < TD->getPointerSizeInBits()) {
10649 if (Constant *C = dyn_cast<Constant>(Op)) {
10650 *i = ConstantExpr::getSExt(C, TD->getIntPtrType());
10653 Op = InsertCastBefore(Instruction::SExt, Op, TD->getIntPtrType(),
10661 if (MadeChange) return &GEP;
10663 // Combine Indices - If the source pointer to this getelementptr instruction
10664 // is a getelementptr instruction, combine the indices of the two
10665 // getelementptr instructions into a single instruction.
10667 SmallVector<Value*, 8> SrcGEPOperands;
10668 if (User *Src = dyn_castGetElementPtr(PtrOp))
10669 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
10671 if (!SrcGEPOperands.empty()) {
10672 // Note that if our source is a gep chain itself that we wait for that
10673 // chain to be resolved before we perform this transformation. This
10674 // avoids us creating a TON of code in some cases.
10676 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
10677 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
10678 return 0; // Wait until our source is folded to completion.
10680 SmallVector<Value*, 8> Indices;
10682 // Find out whether the last index in the source GEP is a sequential idx.
10683 bool EndsWithSequential = false;
10684 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
10685 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
10686 EndsWithSequential = !isa<StructType>(*I);
10688 // Can we combine the two pointer arithmetics offsets?
10689 if (EndsWithSequential) {
10690 // Replace: gep (gep %P, long B), long A, ...
10691 // With: T = long A+B; gep %P, T, ...
10693 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
10694 if (SO1 == Constant::getNullValue(SO1->getType())) {
10696 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
10699 // If they aren't the same type, convert both to an integer of the
10700 // target's pointer size.
10701 if (SO1->getType() != GO1->getType()) {
10702 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
10703 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
10704 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
10705 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
10707 unsigned PS = TD->getPointerSizeInBits();
10708 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
10709 // Convert GO1 to SO1's type.
10710 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
10712 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
10713 // Convert SO1 to GO1's type.
10714 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
10716 const Type *PT = TD->getIntPtrType();
10717 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
10718 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
10722 if (isa<Constant>(SO1) && isa<Constant>(GO1))
10723 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
10725 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
10726 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
10730 // Recycle the GEP we already have if possible.
10731 if (SrcGEPOperands.size() == 2) {
10732 GEP.setOperand(0, SrcGEPOperands[0]);
10733 GEP.setOperand(1, Sum);
10736 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10737 SrcGEPOperands.end()-1);
10738 Indices.push_back(Sum);
10739 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
10741 } else if (isa<Constant>(*GEP.idx_begin()) &&
10742 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
10743 SrcGEPOperands.size() != 1) {
10744 // Otherwise we can do the fold if the first index of the GEP is a zero
10745 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
10746 SrcGEPOperands.end());
10747 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
10750 if (!Indices.empty())
10751 return GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
10752 Indices.end(), GEP.getName());
10754 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
10755 // GEP of global variable. If all of the indices for this GEP are
10756 // constants, we can promote this to a constexpr instead of an instruction.
10758 // Scan for nonconstants...
10759 SmallVector<Constant*, 8> Indices;
10760 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
10761 for (; I != E && isa<Constant>(*I); ++I)
10762 Indices.push_back(cast<Constant>(*I));
10764 if (I == E) { // If they are all constants...
10765 Constant *CE = ConstantExpr::getGetElementPtr(GV,
10766 &Indices[0],Indices.size());
10768 // Replace all uses of the GEP with the new constexpr...
10769 return ReplaceInstUsesWith(GEP, CE);
10771 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
10772 if (!isa<PointerType>(X->getType())) {
10773 // Not interesting. Source pointer must be a cast from pointer.
10774 } else if (HasZeroPointerIndex) {
10775 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
10776 // into : GEP [10 x i8]* X, i32 0, ...
10778 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
10779 // into : GEP i8* X, ...
10781 // This occurs when the program declares an array extern like "int X[];"
10782 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
10783 const PointerType *XTy = cast<PointerType>(X->getType());
10784 if (const ArrayType *CATy =
10785 dyn_cast<ArrayType>(CPTy->getElementType())) {
10786 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
10787 if (CATy->getElementType() == XTy->getElementType()) {
10788 // -> GEP i8* X, ...
10789 SmallVector<Value*, 8> Indices(GEP.idx_begin()+1, GEP.idx_end());
10790 return GetElementPtrInst::Create(X, Indices.begin(), Indices.end(),
10792 } else if (const ArrayType *XATy =
10793 dyn_cast<ArrayType>(XTy->getElementType())) {
10794 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
10795 if (CATy->getElementType() == XATy->getElementType()) {
10796 // -> GEP [10 x i8]* X, i32 0, ...
10797 // At this point, we know that the cast source type is a pointer
10798 // to an array of the same type as the destination pointer
10799 // array. Because the array type is never stepped over (there
10800 // is a leading zero) we can fold the cast into this GEP.
10801 GEP.setOperand(0, X);
10806 } else if (GEP.getNumOperands() == 2) {
10807 // Transform things like:
10808 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
10809 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
10810 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
10811 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
10812 if (isa<ArrayType>(SrcElTy) &&
10813 TD->getTypePaddedSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
10814 TD->getTypePaddedSize(ResElTy)) {
10816 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10817 Idx[1] = GEP.getOperand(1);
10818 Value *V = InsertNewInstBefore(
10819 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName()), GEP);
10820 // V and GEP are both pointer types --> BitCast
10821 return new BitCastInst(V, GEP.getType());
10824 // Transform things like:
10825 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
10826 // (where tmp = 8*tmp2) into:
10827 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
10829 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
10830 uint64_t ArrayEltSize =
10831 TD->getTypePaddedSize(cast<ArrayType>(SrcElTy)->getElementType());
10833 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
10834 // allow either a mul, shift, or constant here.
10836 ConstantInt *Scale = 0;
10837 if (ArrayEltSize == 1) {
10838 NewIdx = GEP.getOperand(1);
10839 Scale = ConstantInt::get(NewIdx->getType(), 1);
10840 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
10841 NewIdx = ConstantInt::get(CI->getType(), 1);
10843 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
10844 if (Inst->getOpcode() == Instruction::Shl &&
10845 isa<ConstantInt>(Inst->getOperand(1))) {
10846 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
10847 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
10848 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
10849 NewIdx = Inst->getOperand(0);
10850 } else if (Inst->getOpcode() == Instruction::Mul &&
10851 isa<ConstantInt>(Inst->getOperand(1))) {
10852 Scale = cast<ConstantInt>(Inst->getOperand(1));
10853 NewIdx = Inst->getOperand(0);
10857 // If the index will be to exactly the right offset with the scale taken
10858 // out, perform the transformation. Note, we don't know whether Scale is
10859 // signed or not. We'll use unsigned version of division/modulo
10860 // operation after making sure Scale doesn't have the sign bit set.
10861 if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL &&
10862 Scale->getZExtValue() % ArrayEltSize == 0) {
10863 Scale = ConstantInt::get(Scale->getType(),
10864 Scale->getZExtValue() / ArrayEltSize);
10865 if (Scale->getZExtValue() != 1) {
10866 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
10868 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
10869 NewIdx = InsertNewInstBefore(Sc, GEP);
10872 // Insert the new GEP instruction.
10874 Idx[0] = Constant::getNullValue(Type::Int32Ty);
10876 Instruction *NewGEP =
10877 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
10878 NewGEP = InsertNewInstBefore(NewGEP, GEP);
10879 // The NewGEP must be pointer typed, so must the old one -> BitCast
10880 return new BitCastInst(NewGEP, GEP.getType());
10886 /// See if we can simplify:
10887 /// X = bitcast A to B*
10888 /// Y = gep X, <...constant indices...>
10889 /// into a gep of the original struct. This is important for SROA and alias
10890 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
10891 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
10892 if (!isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices()) {
10893 // Determine how much the GEP moves the pointer. We are guaranteed to get
10894 // a constant back from EmitGEPOffset.
10895 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(&GEP, GEP, *this));
10896 int64_t Offset = OffsetV->getSExtValue();
10898 // If this GEP instruction doesn't move the pointer, just replace the GEP
10899 // with a bitcast of the real input to the dest type.
10901 // If the bitcast is of an allocation, and the allocation will be
10902 // converted to match the type of the cast, don't touch this.
10903 if (isa<AllocationInst>(BCI->getOperand(0))) {
10904 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
10905 if (Instruction *I = visitBitCast(*BCI)) {
10908 BCI->getParent()->getInstList().insert(BCI, I);
10909 ReplaceInstUsesWith(*BCI, I);
10914 return new BitCastInst(BCI->getOperand(0), GEP.getType());
10917 // Otherwise, if the offset is non-zero, we need to find out if there is a
10918 // field at Offset in 'A's type. If so, we can pull the cast through the
10920 SmallVector<Value*, 8> NewIndices;
10922 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
10923 if (FindElementAtOffset(InTy, Offset, NewIndices, TD)) {
10924 Instruction *NGEP =
10925 GetElementPtrInst::Create(BCI->getOperand(0), NewIndices.begin(),
10927 if (NGEP->getType() == GEP.getType()) return NGEP;
10928 InsertNewInstBefore(NGEP, GEP);
10929 NGEP->takeName(&GEP);
10930 return new BitCastInst(NGEP, GEP.getType());
10938 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
10939 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
10940 if (AI.isArrayAllocation()) { // Check C != 1
10941 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
10942 const Type *NewTy =
10943 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
10944 AllocationInst *New = 0;
10946 // Create and insert the replacement instruction...
10947 if (isa<MallocInst>(AI))
10948 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
10950 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
10951 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
10954 InsertNewInstBefore(New, AI);
10956 // Scan to the end of the allocation instructions, to skip over a block of
10957 // allocas if possible...also skip interleaved debug info
10959 BasicBlock::iterator It = New;
10960 while (isa<AllocationInst>(*It) || isa<DbgInfoIntrinsic>(*It)) ++It;
10962 // Now that I is pointing to the first non-allocation-inst in the block,
10963 // insert our getelementptr instruction...
10965 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
10969 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
10970 New->getName()+".sub", It);
10972 // Now make everything use the getelementptr instead of the original
10974 return ReplaceInstUsesWith(AI, V);
10975 } else if (isa<UndefValue>(AI.getArraySize())) {
10976 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
10980 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized()) {
10981 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
10982 // Note that we only do this for alloca's, because malloc should allocate and
10983 // return a unique pointer, even for a zero byte allocation.
10984 if (TD->getTypePaddedSize(AI.getAllocatedType()) == 0)
10985 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
10987 // If the alignment is 0 (unspecified), assign it the preferred alignment.
10988 if (AI.getAlignment() == 0)
10989 AI.setAlignment(TD->getPrefTypeAlignment(AI.getAllocatedType()));
10995 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
10996 Value *Op = FI.getOperand(0);
10998 // free undef -> unreachable.
10999 if (isa<UndefValue>(Op)) {
11000 // Insert a new store to null because we cannot modify the CFG here.
11001 new StoreInst(ConstantInt::getTrue(),
11002 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
11003 return EraseInstFromFunction(FI);
11006 // If we have 'free null' delete the instruction. This can happen in stl code
11007 // when lots of inlining happens.
11008 if (isa<ConstantPointerNull>(Op))
11009 return EraseInstFromFunction(FI);
11011 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
11012 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
11013 FI.setOperand(0, CI->getOperand(0));
11017 // Change free (gep X, 0,0,0,0) into free(X)
11018 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11019 if (GEPI->hasAllZeroIndices()) {
11020 AddToWorkList(GEPI);
11021 FI.setOperand(0, GEPI->getOperand(0));
11026 // Change free(malloc) into nothing, if the malloc has a single use.
11027 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
11028 if (MI->hasOneUse()) {
11029 EraseInstFromFunction(FI);
11030 return EraseInstFromFunction(*MI);
11037 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
11038 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
11039 const TargetData *TD) {
11040 User *CI = cast<User>(LI.getOperand(0));
11041 Value *CastOp = CI->getOperand(0);
11043 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
11044 // Instead of loading constant c string, use corresponding integer value
11045 // directly if string length is small enough.
11047 if (GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
11048 unsigned len = Str.length();
11049 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
11050 unsigned numBits = Ty->getPrimitiveSizeInBits();
11051 // Replace LI with immediate integer store.
11052 if ((numBits >> 3) == len + 1) {
11053 APInt StrVal(numBits, 0);
11054 APInt SingleChar(numBits, 0);
11055 if (TD->isLittleEndian()) {
11056 for (signed i = len-1; i >= 0; i--) {
11057 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
11058 StrVal = (StrVal << 8) | SingleChar;
11061 for (unsigned i = 0; i < len; i++) {
11062 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
11063 StrVal = (StrVal << 8) | SingleChar;
11065 // Append NULL at the end.
11067 StrVal = (StrVal << 8) | SingleChar;
11069 Value *NL = ConstantInt::get(StrVal);
11070 return IC.ReplaceInstUsesWith(LI, NL);
11075 const PointerType *DestTy = cast<PointerType>(CI->getType());
11076 const Type *DestPTy = DestTy->getElementType();
11077 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
11079 // If the address spaces don't match, don't eliminate the cast.
11080 if (DestTy->getAddressSpace() != SrcTy->getAddressSpace())
11083 const Type *SrcPTy = SrcTy->getElementType();
11085 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
11086 isa<VectorType>(DestPTy)) {
11087 // If the source is an array, the code below will not succeed. Check to
11088 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11090 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
11091 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
11092 if (ASrcTy->getNumElements() != 0) {
11094 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
11095 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
11096 SrcTy = cast<PointerType>(CastOp->getType());
11097 SrcPTy = SrcTy->getElementType();
11100 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
11101 isa<VectorType>(SrcPTy)) &&
11102 // Do not allow turning this into a load of an integer, which is then
11103 // casted to a pointer, this pessimizes pointer analysis a lot.
11104 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
11105 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
11106 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
11108 // Okay, we are casting from one integer or pointer type to another of
11109 // the same size. Instead of casting the pointer before the load, cast
11110 // the result of the loaded value.
11111 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
11113 LI.isVolatile()),LI);
11114 // Now cast the result of the load.
11115 return new BitCastInst(NewLoad, LI.getType());
11122 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
11123 /// from this value cannot trap. If it is not obviously safe to load from the
11124 /// specified pointer, we do a quick local scan of the basic block containing
11125 /// ScanFrom, to determine if the address is already accessed.
11126 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
11127 // If it is an alloca it is always safe to load from.
11128 if (isa<AllocaInst>(V)) return true;
11130 // If it is a global variable it is mostly safe to load from.
11131 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
11132 // Don't try to evaluate aliases. External weak GV can be null.
11133 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
11135 // Otherwise, be a little bit agressive by scanning the local block where we
11136 // want to check to see if the pointer is already being loaded or stored
11137 // from/to. If so, the previous load or store would have already trapped,
11138 // so there is no harm doing an extra load (also, CSE will later eliminate
11139 // the load entirely).
11140 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
11145 // If we see a free or a call (which might do a free) the pointer could be
11147 if (isa<FreeInst>(BBI) || isa<CallInst>(BBI))
11150 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11151 if (LI->getOperand(0) == V) return true;
11152 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) {
11153 if (SI->getOperand(1) == V) return true;
11160 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
11161 Value *Op = LI.getOperand(0);
11163 // Attempt to improve the alignment.
11164 unsigned KnownAlign =
11165 GetOrEnforceKnownAlignment(Op, TD->getPrefTypeAlignment(LI.getType()));
11167 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
11168 LI.getAlignment()))
11169 LI.setAlignment(KnownAlign);
11171 // load (cast X) --> cast (load X) iff safe
11172 if (isa<CastInst>(Op))
11173 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11176 // None of the following transforms are legal for volatile loads.
11177 if (LI.isVolatile()) return 0;
11179 // Do really simple store-to-load forwarding and load CSE, to catch cases
11180 // where there are several consequtive memory accesses to the same location,
11181 // separated by a few arithmetic operations.
11182 BasicBlock::iterator BBI = &LI;
11183 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
11184 return ReplaceInstUsesWith(LI, AvailableVal);
11186 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11187 const Value *GEPI0 = GEPI->getOperand(0);
11188 // TODO: Consider a target hook for valid address spaces for this xform.
11189 if (isa<ConstantPointerNull>(GEPI0) &&
11190 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
11191 // Insert a new store to null instruction before the load to indicate
11192 // that this code is not reachable. We do this instead of inserting
11193 // an unreachable instruction directly because we cannot modify the
11195 new StoreInst(UndefValue::get(LI.getType()),
11196 Constant::getNullValue(Op->getType()), &LI);
11197 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11201 if (Constant *C = dyn_cast<Constant>(Op)) {
11202 // load null/undef -> undef
11203 // TODO: Consider a target hook for valid address spaces for this xform.
11204 if (isa<UndefValue>(C) || (C->isNullValue() &&
11205 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
11206 // Insert a new store to null instruction before the load to indicate that
11207 // this code is not reachable. We do this instead of inserting an
11208 // unreachable instruction directly because we cannot modify the CFG.
11209 new StoreInst(UndefValue::get(LI.getType()),
11210 Constant::getNullValue(Op->getType()), &LI);
11211 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11214 // Instcombine load (constant global) into the value loaded.
11215 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
11216 if (GV->isConstant() && !GV->isDeclaration())
11217 return ReplaceInstUsesWith(LI, GV->getInitializer());
11219 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
11220 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
11221 if (CE->getOpcode() == Instruction::GetElementPtr) {
11222 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
11223 if (GV->isConstant() && !GV->isDeclaration())
11225 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
11226 return ReplaceInstUsesWith(LI, V);
11227 if (CE->getOperand(0)->isNullValue()) {
11228 // Insert a new store to null instruction before the load to indicate
11229 // that this code is not reachable. We do this instead of inserting
11230 // an unreachable instruction directly because we cannot modify the
11232 new StoreInst(UndefValue::get(LI.getType()),
11233 Constant::getNullValue(Op->getType()), &LI);
11234 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11237 } else if (CE->isCast()) {
11238 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11244 // If this load comes from anywhere in a constant global, and if the global
11245 // is all undef or zero, we know what it loads.
11246 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op->getUnderlyingObject())){
11247 if (GV->isConstant() && GV->hasInitializer()) {
11248 if (GV->getInitializer()->isNullValue())
11249 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
11250 else if (isa<UndefValue>(GV->getInitializer()))
11251 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11255 if (Op->hasOneUse()) {
11256 // Change select and PHI nodes to select values instead of addresses: this
11257 // helps alias analysis out a lot, allows many others simplifications, and
11258 // exposes redundancy in the code.
11260 // Note that we cannot do the transformation unless we know that the
11261 // introduced loads cannot trap! Something like this is valid as long as
11262 // the condition is always false: load (select bool %C, int* null, int* %G),
11263 // but it would not be valid if we transformed it to load from null
11264 // unconditionally.
11266 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
11267 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
11268 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
11269 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
11270 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
11271 SI->getOperand(1)->getName()+".val"), LI);
11272 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
11273 SI->getOperand(2)->getName()+".val"), LI);
11274 return SelectInst::Create(SI->getCondition(), V1, V2);
11277 // load (select (cond, null, P)) -> load P
11278 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
11279 if (C->isNullValue()) {
11280 LI.setOperand(0, SI->getOperand(2));
11284 // load (select (cond, P, null)) -> load P
11285 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
11286 if (C->isNullValue()) {
11287 LI.setOperand(0, SI->getOperand(1));
11295 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
11296 /// when possible. This makes it generally easy to do alias analysis and/or
11297 /// SROA/mem2reg of the memory object.
11298 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
11299 User *CI = cast<User>(SI.getOperand(1));
11300 Value *CastOp = CI->getOperand(0);
11302 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
11303 const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType());
11304 if (SrcTy == 0) return 0;
11306 const Type *SrcPTy = SrcTy->getElementType();
11308 if (!DestPTy->isInteger() && !isa<PointerType>(DestPTy))
11311 /// NewGEPIndices - If SrcPTy is an aggregate type, we can emit a "noop gep"
11312 /// to its first element. This allows us to handle things like:
11313 /// store i32 xxx, (bitcast {foo*, float}* %P to i32*)
11314 /// on 32-bit hosts.
11315 SmallVector<Value*, 4> NewGEPIndices;
11317 // If the source is an array, the code below will not succeed. Check to
11318 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11320 if (isa<ArrayType>(SrcPTy) || isa<StructType>(SrcPTy)) {
11321 // Index through pointer.
11322 Constant *Zero = Constant::getNullValue(Type::Int32Ty);
11323 NewGEPIndices.push_back(Zero);
11326 if (const StructType *STy = dyn_cast<StructType>(SrcPTy)) {
11327 if (!STy->getNumElements()) /* Struct can be empty {} */
11329 NewGEPIndices.push_back(Zero);
11330 SrcPTy = STy->getElementType(0);
11331 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcPTy)) {
11332 NewGEPIndices.push_back(Zero);
11333 SrcPTy = ATy->getElementType();
11339 SrcTy = PointerType::get(SrcPTy, SrcTy->getAddressSpace());
11342 if (!SrcPTy->isInteger() && !isa<PointerType>(SrcPTy))
11345 // If the pointers point into different address spaces or if they point to
11346 // values with different sizes, we can't do the transformation.
11347 if (SrcTy->getAddressSpace() !=
11348 cast<PointerType>(CI->getType())->getAddressSpace() ||
11349 IC.getTargetData().getTypeSizeInBits(SrcPTy) !=
11350 IC.getTargetData().getTypeSizeInBits(DestPTy))
11353 // Okay, we are casting from one integer or pointer type to another of
11354 // the same size. Instead of casting the pointer before
11355 // the store, cast the value to be stored.
11357 Value *SIOp0 = SI.getOperand(0);
11358 Instruction::CastOps opcode = Instruction::BitCast;
11359 const Type* CastSrcTy = SIOp0->getType();
11360 const Type* CastDstTy = SrcPTy;
11361 if (isa<PointerType>(CastDstTy)) {
11362 if (CastSrcTy->isInteger())
11363 opcode = Instruction::IntToPtr;
11364 } else if (isa<IntegerType>(CastDstTy)) {
11365 if (isa<PointerType>(SIOp0->getType()))
11366 opcode = Instruction::PtrToInt;
11369 // SIOp0 is a pointer to aggregate and this is a store to the first field,
11370 // emit a GEP to index into its first field.
11371 if (!NewGEPIndices.empty()) {
11372 if (Constant *C = dyn_cast<Constant>(CastOp))
11373 CastOp = ConstantExpr::getGetElementPtr(C, &NewGEPIndices[0],
11374 NewGEPIndices.size());
11376 CastOp = IC.InsertNewInstBefore(
11377 GetElementPtrInst::Create(CastOp, NewGEPIndices.begin(),
11378 NewGEPIndices.end()), SI);
11381 if (Constant *C = dyn_cast<Constant>(SIOp0))
11382 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
11384 NewCast = IC.InsertNewInstBefore(
11385 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
11387 return new StoreInst(NewCast, CastOp);
11390 /// equivalentAddressValues - Test if A and B will obviously have the same
11391 /// value. This includes recognizing that %t0 and %t1 will have the same
11392 /// value in code like this:
11393 /// %t0 = getelementptr \@a, 0, 3
11394 /// store i32 0, i32* %t0
11395 /// %t1 = getelementptr \@a, 0, 3
11396 /// %t2 = load i32* %t1
11398 static bool equivalentAddressValues(Value *A, Value *B) {
11399 // Test if the values are trivially equivalent.
11400 if (A == B) return true;
11402 // Test if the values come form identical arithmetic instructions.
11403 if (isa<BinaryOperator>(A) ||
11404 isa<CastInst>(A) ||
11406 isa<GetElementPtrInst>(A))
11407 if (Instruction *BI = dyn_cast<Instruction>(B))
11408 if (cast<Instruction>(A)->isIdenticalTo(BI))
11411 // Otherwise they may not be equivalent.
11415 // If this instruction has two uses, one of which is a llvm.dbg.declare,
11416 // return the llvm.dbg.declare.
11417 DbgDeclareInst *InstCombiner::hasOneUsePlusDeclare(Value *V) {
11418 if (!V->hasNUses(2))
11420 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
11422 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI))
11424 if (isa<BitCastInst>(UI) && UI->hasOneUse()) {
11425 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI->use_begin()))
11432 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
11433 Value *Val = SI.getOperand(0);
11434 Value *Ptr = SI.getOperand(1);
11436 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
11437 EraseInstFromFunction(SI);
11442 // If the RHS is an alloca with a single use, zapify the store, making the
11444 // If the RHS is an alloca with a two uses, the other one being a
11445 // llvm.dbg.declare, zapify the store and the declare, making the
11446 // alloca dead. We must do this to prevent declare's from affecting
11448 if (!SI.isVolatile()) {
11449 if (Ptr->hasOneUse()) {
11450 if (isa<AllocaInst>(Ptr)) {
11451 EraseInstFromFunction(SI);
11455 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
11456 if (isa<AllocaInst>(GEP->getOperand(0))) {
11457 if (GEP->getOperand(0)->hasOneUse()) {
11458 EraseInstFromFunction(SI);
11462 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(GEP->getOperand(0))) {
11463 EraseInstFromFunction(*DI);
11464 EraseInstFromFunction(SI);
11471 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(Ptr)) {
11472 EraseInstFromFunction(*DI);
11473 EraseInstFromFunction(SI);
11479 // Attempt to improve the alignment.
11480 unsigned KnownAlign =
11481 GetOrEnforceKnownAlignment(Ptr, TD->getPrefTypeAlignment(Val->getType()));
11483 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
11484 SI.getAlignment()))
11485 SI.setAlignment(KnownAlign);
11487 // Do really simple DSE, to catch cases where there are several consecutive
11488 // stores to the same location, separated by a few arithmetic operations. This
11489 // situation often occurs with bitfield accesses.
11490 BasicBlock::iterator BBI = &SI;
11491 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
11494 // Don't count debug info directives, lest they affect codegen,
11495 // and we skip pointer-to-pointer bitcasts, which are NOPs.
11496 // It is necessary for correctness to skip those that feed into a
11497 // llvm.dbg.declare, as these are not present when debugging is off.
11498 if (isa<DbgInfoIntrinsic>(BBI) ||
11499 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
11504 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
11505 // Prev store isn't volatile, and stores to the same location?
11506 if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
11507 SI.getOperand(1))) {
11510 EraseInstFromFunction(*PrevSI);
11516 // If this is a load, we have to stop. However, if the loaded value is from
11517 // the pointer we're loading and is producing the pointer we're storing,
11518 // then *this* store is dead (X = load P; store X -> P).
11519 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11520 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
11521 !SI.isVolatile()) {
11522 EraseInstFromFunction(SI);
11526 // Otherwise, this is a load from some other location. Stores before it
11527 // may not be dead.
11531 // Don't skip over loads or things that can modify memory.
11532 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
11537 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
11539 // store X, null -> turns into 'unreachable' in SimplifyCFG
11540 if (isa<ConstantPointerNull>(Ptr)) {
11541 if (!isa<UndefValue>(Val)) {
11542 SI.setOperand(0, UndefValue::get(Val->getType()));
11543 if (Instruction *U = dyn_cast<Instruction>(Val))
11544 AddToWorkList(U); // Dropped a use.
11547 return 0; // Do not modify these!
11550 // store undef, Ptr -> noop
11551 if (isa<UndefValue>(Val)) {
11552 EraseInstFromFunction(SI);
11557 // If the pointer destination is a cast, see if we can fold the cast into the
11559 if (isa<CastInst>(Ptr))
11560 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11562 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
11564 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11568 // If this store is the last instruction in the basic block (possibly
11569 // excepting debug info instructions and the pointer bitcasts that feed
11570 // into them), and if the block ends with an unconditional branch, try
11571 // to move it to the successor block.
11575 } while (isa<DbgInfoIntrinsic>(BBI) ||
11576 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType())));
11577 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
11578 if (BI->isUnconditional())
11579 if (SimplifyStoreAtEndOfBlock(SI))
11580 return 0; // xform done!
11585 /// SimplifyStoreAtEndOfBlock - Turn things like:
11586 /// if () { *P = v1; } else { *P = v2 }
11587 /// into a phi node with a store in the successor.
11589 /// Simplify things like:
11590 /// *P = v1; if () { *P = v2; }
11591 /// into a phi node with a store in the successor.
11593 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
11594 BasicBlock *StoreBB = SI.getParent();
11596 // Check to see if the successor block has exactly two incoming edges. If
11597 // so, see if the other predecessor contains a store to the same location.
11598 // if so, insert a PHI node (if needed) and move the stores down.
11599 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
11601 // Determine whether Dest has exactly two predecessors and, if so, compute
11602 // the other predecessor.
11603 pred_iterator PI = pred_begin(DestBB);
11604 BasicBlock *OtherBB = 0;
11605 if (*PI != StoreBB)
11608 if (PI == pred_end(DestBB))
11611 if (*PI != StoreBB) {
11616 if (++PI != pred_end(DestBB))
11619 // Bail out if all the relevant blocks aren't distinct (this can happen,
11620 // for example, if SI is in an infinite loop)
11621 if (StoreBB == DestBB || OtherBB == DestBB)
11624 // Verify that the other block ends in a branch and is not otherwise empty.
11625 BasicBlock::iterator BBI = OtherBB->getTerminator();
11626 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
11627 if (!OtherBr || BBI == OtherBB->begin())
11630 // If the other block ends in an unconditional branch, check for the 'if then
11631 // else' case. there is an instruction before the branch.
11632 StoreInst *OtherStore = 0;
11633 if (OtherBr->isUnconditional()) {
11635 // Skip over debugging info.
11636 while (isa<DbgInfoIntrinsic>(BBI) ||
11637 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
11638 if (BBI==OtherBB->begin())
11642 // If this isn't a store, or isn't a store to the same location, bail out.
11643 OtherStore = dyn_cast<StoreInst>(BBI);
11644 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
11647 // Otherwise, the other block ended with a conditional branch. If one of the
11648 // destinations is StoreBB, then we have the if/then case.
11649 if (OtherBr->getSuccessor(0) != StoreBB &&
11650 OtherBr->getSuccessor(1) != StoreBB)
11653 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
11654 // if/then triangle. See if there is a store to the same ptr as SI that
11655 // lives in OtherBB.
11657 // Check to see if we find the matching store.
11658 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
11659 if (OtherStore->getOperand(1) != SI.getOperand(1))
11663 // If we find something that may be using or overwriting the stored
11664 // value, or if we run out of instructions, we can't do the xform.
11665 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
11666 BBI == OtherBB->begin())
11670 // In order to eliminate the store in OtherBr, we have to
11671 // make sure nothing reads or overwrites the stored value in
11673 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
11674 // FIXME: This should really be AA driven.
11675 if (I->mayReadFromMemory() || I->mayWriteToMemory())
11680 // Insert a PHI node now if we need it.
11681 Value *MergedVal = OtherStore->getOperand(0);
11682 if (MergedVal != SI.getOperand(0)) {
11683 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
11684 PN->reserveOperandSpace(2);
11685 PN->addIncoming(SI.getOperand(0), SI.getParent());
11686 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
11687 MergedVal = InsertNewInstBefore(PN, DestBB->front());
11690 // Advance to a place where it is safe to insert the new store and
11692 BBI = DestBB->getFirstNonPHI();
11693 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
11694 OtherStore->isVolatile()), *BBI);
11696 // Nuke the old stores.
11697 EraseInstFromFunction(SI);
11698 EraseInstFromFunction(*OtherStore);
11704 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
11705 // Change br (not X), label True, label False to: br X, label False, True
11707 BasicBlock *TrueDest;
11708 BasicBlock *FalseDest;
11709 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
11710 !isa<Constant>(X)) {
11711 // Swap Destinations and condition...
11712 BI.setCondition(X);
11713 BI.setSuccessor(0, FalseDest);
11714 BI.setSuccessor(1, TrueDest);
11718 // Cannonicalize fcmp_one -> fcmp_oeq
11719 FCmpInst::Predicate FPred; Value *Y;
11720 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
11721 TrueDest, FalseDest)))
11722 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
11723 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
11724 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
11725 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
11726 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
11727 NewSCC->takeName(I);
11728 // Swap Destinations and condition...
11729 BI.setCondition(NewSCC);
11730 BI.setSuccessor(0, FalseDest);
11731 BI.setSuccessor(1, TrueDest);
11732 RemoveFromWorkList(I);
11733 I->eraseFromParent();
11734 AddToWorkList(NewSCC);
11738 // Cannonicalize icmp_ne -> icmp_eq
11739 ICmpInst::Predicate IPred;
11740 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
11741 TrueDest, FalseDest)))
11742 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
11743 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
11744 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
11745 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
11746 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
11747 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
11748 NewSCC->takeName(I);
11749 // Swap Destinations and condition...
11750 BI.setCondition(NewSCC);
11751 BI.setSuccessor(0, FalseDest);
11752 BI.setSuccessor(1, TrueDest);
11753 RemoveFromWorkList(I);
11754 I->eraseFromParent();;
11755 AddToWorkList(NewSCC);
11762 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
11763 Value *Cond = SI.getCondition();
11764 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
11765 if (I->getOpcode() == Instruction::Add)
11766 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
11767 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
11768 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
11769 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
11771 SI.setOperand(0, I->getOperand(0));
11779 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
11780 Value *Agg = EV.getAggregateOperand();
11782 if (!EV.hasIndices())
11783 return ReplaceInstUsesWith(EV, Agg);
11785 if (Constant *C = dyn_cast<Constant>(Agg)) {
11786 if (isa<UndefValue>(C))
11787 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
11789 if (isa<ConstantAggregateZero>(C))
11790 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
11792 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
11793 // Extract the element indexed by the first index out of the constant
11794 Value *V = C->getOperand(*EV.idx_begin());
11795 if (EV.getNumIndices() > 1)
11796 // Extract the remaining indices out of the constant indexed by the
11798 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
11800 return ReplaceInstUsesWith(EV, V);
11802 return 0; // Can't handle other constants
11804 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
11805 // We're extracting from an insertvalue instruction, compare the indices
11806 const unsigned *exti, *exte, *insi, *inse;
11807 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
11808 exte = EV.idx_end(), inse = IV->idx_end();
11809 exti != exte && insi != inse;
11811 if (*insi != *exti)
11812 // The insert and extract both reference distinctly different elements.
11813 // This means the extract is not influenced by the insert, and we can
11814 // replace the aggregate operand of the extract with the aggregate
11815 // operand of the insert. i.e., replace
11816 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
11817 // %E = extractvalue { i32, { i32 } } %I, 0
11819 // %E = extractvalue { i32, { i32 } } %A, 0
11820 return ExtractValueInst::Create(IV->getAggregateOperand(),
11821 EV.idx_begin(), EV.idx_end());
11823 if (exti == exte && insi == inse)
11824 // Both iterators are at the end: Index lists are identical. Replace
11825 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
11826 // %C = extractvalue { i32, { i32 } } %B, 1, 0
11828 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
11829 if (exti == exte) {
11830 // The extract list is a prefix of the insert list. i.e. replace
11831 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
11832 // %E = extractvalue { i32, { i32 } } %I, 1
11834 // %X = extractvalue { i32, { i32 } } %A, 1
11835 // %E = insertvalue { i32 } %X, i32 42, 0
11836 // by switching the order of the insert and extract (though the
11837 // insertvalue should be left in, since it may have other uses).
11838 Value *NewEV = InsertNewInstBefore(
11839 ExtractValueInst::Create(IV->getAggregateOperand(),
11840 EV.idx_begin(), EV.idx_end()),
11842 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
11846 // The insert list is a prefix of the extract list
11847 // We can simply remove the common indices from the extract and make it
11848 // operate on the inserted value instead of the insertvalue result.
11850 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
11851 // %E = extractvalue { i32, { i32 } } %I, 1, 0
11853 // %E extractvalue { i32 } { i32 42 }, 0
11854 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
11857 // Can't simplify extracts from other values. Note that nested extracts are
11858 // already simplified implicitely by the above (extract ( extract (insert) )
11859 // will be translated into extract ( insert ( extract ) ) first and then just
11860 // the value inserted, if appropriate).
11864 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
11865 /// is to leave as a vector operation.
11866 static bool CheapToScalarize(Value *V, bool isConstant) {
11867 if (isa<ConstantAggregateZero>(V))
11869 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
11870 if (isConstant) return true;
11871 // If all elts are the same, we can extract.
11872 Constant *Op0 = C->getOperand(0);
11873 for (unsigned i = 1; i < C->getNumOperands(); ++i)
11874 if (C->getOperand(i) != Op0)
11878 Instruction *I = dyn_cast<Instruction>(V);
11879 if (!I) return false;
11881 // Insert element gets simplified to the inserted element or is deleted if
11882 // this is constant idx extract element and its a constant idx insertelt.
11883 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
11884 isa<ConstantInt>(I->getOperand(2)))
11886 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
11888 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
11889 if (BO->hasOneUse() &&
11890 (CheapToScalarize(BO->getOperand(0), isConstant) ||
11891 CheapToScalarize(BO->getOperand(1), isConstant)))
11893 if (CmpInst *CI = dyn_cast<CmpInst>(I))
11894 if (CI->hasOneUse() &&
11895 (CheapToScalarize(CI->getOperand(0), isConstant) ||
11896 CheapToScalarize(CI->getOperand(1), isConstant)))
11902 /// Read and decode a shufflevector mask.
11904 /// It turns undef elements into values that are larger than the number of
11905 /// elements in the input.
11906 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
11907 unsigned NElts = SVI->getType()->getNumElements();
11908 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
11909 return std::vector<unsigned>(NElts, 0);
11910 if (isa<UndefValue>(SVI->getOperand(2)))
11911 return std::vector<unsigned>(NElts, 2*NElts);
11913 std::vector<unsigned> Result;
11914 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
11915 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
11916 if (isa<UndefValue>(*i))
11917 Result.push_back(NElts*2); // undef -> 8
11919 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
11923 /// FindScalarElement - Given a vector and an element number, see if the scalar
11924 /// value is already around as a register, for example if it were inserted then
11925 /// extracted from the vector.
11926 static Value *FindScalarElement(Value *V, unsigned EltNo) {
11927 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
11928 const VectorType *PTy = cast<VectorType>(V->getType());
11929 unsigned Width = PTy->getNumElements();
11930 if (EltNo >= Width) // Out of range access.
11931 return UndefValue::get(PTy->getElementType());
11933 if (isa<UndefValue>(V))
11934 return UndefValue::get(PTy->getElementType());
11935 else if (isa<ConstantAggregateZero>(V))
11936 return Constant::getNullValue(PTy->getElementType());
11937 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
11938 return CP->getOperand(EltNo);
11939 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
11940 // If this is an insert to a variable element, we don't know what it is.
11941 if (!isa<ConstantInt>(III->getOperand(2)))
11943 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
11945 // If this is an insert to the element we are looking for, return the
11947 if (EltNo == IIElt)
11948 return III->getOperand(1);
11950 // Otherwise, the insertelement doesn't modify the value, recurse on its
11952 return FindScalarElement(III->getOperand(0), EltNo);
11953 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
11954 unsigned LHSWidth =
11955 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
11956 unsigned InEl = getShuffleMask(SVI)[EltNo];
11957 if (InEl < LHSWidth)
11958 return FindScalarElement(SVI->getOperand(0), InEl);
11959 else if (InEl < LHSWidth*2)
11960 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth);
11962 return UndefValue::get(PTy->getElementType());
11965 // Otherwise, we don't know.
11969 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
11970 // If vector val is undef, replace extract with scalar undef.
11971 if (isa<UndefValue>(EI.getOperand(0)))
11972 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
11974 // If vector val is constant 0, replace extract with scalar 0.
11975 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
11976 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
11978 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
11979 // If vector val is constant with all elements the same, replace EI with
11980 // that element. When the elements are not identical, we cannot replace yet
11981 // (we do that below, but only when the index is constant).
11982 Constant *op0 = C->getOperand(0);
11983 for (unsigned i = 1; i < C->getNumOperands(); ++i)
11984 if (C->getOperand(i) != op0) {
11989 return ReplaceInstUsesWith(EI, op0);
11992 // If extracting a specified index from the vector, see if we can recursively
11993 // find a previously computed scalar that was inserted into the vector.
11994 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
11995 unsigned IndexVal = IdxC->getZExtValue();
11996 unsigned VectorWidth =
11997 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
11999 // If this is extracting an invalid index, turn this into undef, to avoid
12000 // crashing the code below.
12001 if (IndexVal >= VectorWidth)
12002 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12004 // This instruction only demands the single element from the input vector.
12005 // If the input vector has a single use, simplify it based on this use
12007 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
12008 APInt UndefElts(VectorWidth, 0);
12009 APInt DemandedMask(VectorWidth, 1 << IndexVal);
12010 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
12011 DemandedMask, UndefElts)) {
12012 EI.setOperand(0, V);
12017 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
12018 return ReplaceInstUsesWith(EI, Elt);
12020 // If the this extractelement is directly using a bitcast from a vector of
12021 // the same number of elements, see if we can find the source element from
12022 // it. In this case, we will end up needing to bitcast the scalars.
12023 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
12024 if (const VectorType *VT =
12025 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
12026 if (VT->getNumElements() == VectorWidth)
12027 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
12028 return new BitCastInst(Elt, EI.getType());
12032 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
12033 if (I->hasOneUse()) {
12034 // Push extractelement into predecessor operation if legal and
12035 // profitable to do so
12036 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
12037 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
12038 if (CheapToScalarize(BO, isConstantElt)) {
12039 ExtractElementInst *newEI0 =
12040 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
12041 EI.getName()+".lhs");
12042 ExtractElementInst *newEI1 =
12043 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
12044 EI.getName()+".rhs");
12045 InsertNewInstBefore(newEI0, EI);
12046 InsertNewInstBefore(newEI1, EI);
12047 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
12049 } else if (isa<LoadInst>(I)) {
12051 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
12052 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
12053 PointerType::get(EI.getType(), AS),EI);
12054 GetElementPtrInst *GEP =
12055 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
12056 InsertNewInstBefore(GEP, EI);
12057 return new LoadInst(GEP);
12060 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
12061 // Extracting the inserted element?
12062 if (IE->getOperand(2) == EI.getOperand(1))
12063 return ReplaceInstUsesWith(EI, IE->getOperand(1));
12064 // If the inserted and extracted elements are constants, they must not
12065 // be the same value, extract from the pre-inserted value instead.
12066 if (isa<Constant>(IE->getOperand(2)) &&
12067 isa<Constant>(EI.getOperand(1))) {
12068 AddUsesToWorkList(EI);
12069 EI.setOperand(0, IE->getOperand(0));
12072 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
12073 // If this is extracting an element from a shufflevector, figure out where
12074 // it came from and extract from the appropriate input element instead.
12075 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12076 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
12078 unsigned LHSWidth =
12079 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12081 if (SrcIdx < LHSWidth)
12082 Src = SVI->getOperand(0);
12083 else if (SrcIdx < LHSWidth*2) {
12084 SrcIdx -= LHSWidth;
12085 Src = SVI->getOperand(1);
12087 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12089 return new ExtractElementInst(Src, SrcIdx);
12096 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
12097 /// elements from either LHS or RHS, return the shuffle mask and true.
12098 /// Otherwise, return false.
12099 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
12100 std::vector<Constant*> &Mask) {
12101 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
12102 "Invalid CollectSingleShuffleElements");
12103 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12105 if (isa<UndefValue>(V)) {
12106 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
12108 } else if (V == LHS) {
12109 for (unsigned i = 0; i != NumElts; ++i)
12110 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
12112 } else if (V == RHS) {
12113 for (unsigned i = 0; i != NumElts; ++i)
12114 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
12116 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12117 // If this is an insert of an extract from some other vector, include it.
12118 Value *VecOp = IEI->getOperand(0);
12119 Value *ScalarOp = IEI->getOperand(1);
12120 Value *IdxOp = IEI->getOperand(2);
12122 if (!isa<ConstantInt>(IdxOp))
12124 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12126 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
12127 // Okay, we can handle this if the vector we are insertinting into is
12128 // transitively ok.
12129 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
12130 // If so, update the mask to reflect the inserted undef.
12131 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
12134 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
12135 if (isa<ConstantInt>(EI->getOperand(1)) &&
12136 EI->getOperand(0)->getType() == V->getType()) {
12137 unsigned ExtractedIdx =
12138 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12140 // This must be extracting from either LHS or RHS.
12141 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
12142 // Okay, we can handle this if the vector we are insertinting into is
12143 // transitively ok.
12144 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
12145 // If so, update the mask to reflect the inserted value.
12146 if (EI->getOperand(0) == LHS) {
12147 Mask[InsertedIdx % NumElts] =
12148 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
12150 assert(EI->getOperand(0) == RHS);
12151 Mask[InsertedIdx % NumElts] =
12152 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
12161 // TODO: Handle shufflevector here!
12166 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
12167 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
12168 /// that computes V and the LHS value of the shuffle.
12169 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
12171 assert(isa<VectorType>(V->getType()) &&
12172 (RHS == 0 || V->getType() == RHS->getType()) &&
12173 "Invalid shuffle!");
12174 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12176 if (isa<UndefValue>(V)) {
12177 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
12179 } else if (isa<ConstantAggregateZero>(V)) {
12180 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
12182 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12183 // If this is an insert of an extract from some other vector, include it.
12184 Value *VecOp = IEI->getOperand(0);
12185 Value *ScalarOp = IEI->getOperand(1);
12186 Value *IdxOp = IEI->getOperand(2);
12188 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12189 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12190 EI->getOperand(0)->getType() == V->getType()) {
12191 unsigned ExtractedIdx =
12192 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12193 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12195 // Either the extracted from or inserted into vector must be RHSVec,
12196 // otherwise we'd end up with a shuffle of three inputs.
12197 if (EI->getOperand(0) == RHS || RHS == 0) {
12198 RHS = EI->getOperand(0);
12199 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
12200 Mask[InsertedIdx % NumElts] =
12201 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
12205 if (VecOp == RHS) {
12206 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
12207 // Everything but the extracted element is replaced with the RHS.
12208 for (unsigned i = 0; i != NumElts; ++i) {
12209 if (i != InsertedIdx)
12210 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
12215 // If this insertelement is a chain that comes from exactly these two
12216 // vectors, return the vector and the effective shuffle.
12217 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
12218 return EI->getOperand(0);
12223 // TODO: Handle shufflevector here!
12225 // Otherwise, can't do anything fancy. Return an identity vector.
12226 for (unsigned i = 0; i != NumElts; ++i)
12227 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
12231 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
12232 Value *VecOp = IE.getOperand(0);
12233 Value *ScalarOp = IE.getOperand(1);
12234 Value *IdxOp = IE.getOperand(2);
12236 // Inserting an undef or into an undefined place, remove this.
12237 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
12238 ReplaceInstUsesWith(IE, VecOp);
12240 // If the inserted element was extracted from some other vector, and if the
12241 // indexes are constant, try to turn this into a shufflevector operation.
12242 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12243 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12244 EI->getOperand(0)->getType() == IE.getType()) {
12245 unsigned NumVectorElts = IE.getType()->getNumElements();
12246 unsigned ExtractedIdx =
12247 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12248 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12250 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
12251 return ReplaceInstUsesWith(IE, VecOp);
12253 if (InsertedIdx >= NumVectorElts) // Out of range insert.
12254 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
12256 // If we are extracting a value from a vector, then inserting it right
12257 // back into the same place, just use the input vector.
12258 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
12259 return ReplaceInstUsesWith(IE, VecOp);
12261 // We could theoretically do this for ANY input. However, doing so could
12262 // turn chains of insertelement instructions into a chain of shufflevector
12263 // instructions, and right now we do not merge shufflevectors. As such,
12264 // only do this in a situation where it is clear that there is benefit.
12265 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
12266 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
12267 // the values of VecOp, except then one read from EIOp0.
12268 // Build a new shuffle mask.
12269 std::vector<Constant*> Mask;
12270 if (isa<UndefValue>(VecOp))
12271 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
12273 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
12274 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
12277 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
12278 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
12279 ConstantVector::get(Mask));
12282 // If this insertelement isn't used by some other insertelement, turn it
12283 // (and any insertelements it points to), into one big shuffle.
12284 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
12285 std::vector<Constant*> Mask;
12287 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
12288 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
12289 // We now have a shuffle of LHS, RHS, Mask.
12290 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
12299 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
12300 Value *LHS = SVI.getOperand(0);
12301 Value *RHS = SVI.getOperand(1);
12302 std::vector<unsigned> Mask = getShuffleMask(&SVI);
12304 bool MadeChange = false;
12306 // Undefined shuffle mask -> undefined value.
12307 if (isa<UndefValue>(SVI.getOperand(2)))
12308 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
12310 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
12312 if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
12315 APInt UndefElts(VWidth, 0);
12316 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12317 if (SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
12318 LHS = SVI.getOperand(0);
12319 RHS = SVI.getOperand(1);
12323 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
12324 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
12325 if (LHS == RHS || isa<UndefValue>(LHS)) {
12326 if (isa<UndefValue>(LHS) && LHS == RHS) {
12327 // shuffle(undef,undef,mask) -> undef.
12328 return ReplaceInstUsesWith(SVI, LHS);
12331 // Remap any references to RHS to use LHS.
12332 std::vector<Constant*> Elts;
12333 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12334 if (Mask[i] >= 2*e)
12335 Elts.push_back(UndefValue::get(Type::Int32Ty));
12337 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
12338 (Mask[i] < e && isa<UndefValue>(LHS))) {
12339 Mask[i] = 2*e; // Turn into undef.
12340 Elts.push_back(UndefValue::get(Type::Int32Ty));
12342 Mask[i] = Mask[i] % e; // Force to LHS.
12343 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
12347 SVI.setOperand(0, SVI.getOperand(1));
12348 SVI.setOperand(1, UndefValue::get(RHS->getType()));
12349 SVI.setOperand(2, ConstantVector::get(Elts));
12350 LHS = SVI.getOperand(0);
12351 RHS = SVI.getOperand(1);
12355 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
12356 bool isLHSID = true, isRHSID = true;
12358 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12359 if (Mask[i] >= e*2) continue; // Ignore undef values.
12360 // Is this an identity shuffle of the LHS value?
12361 isLHSID &= (Mask[i] == i);
12363 // Is this an identity shuffle of the RHS value?
12364 isRHSID &= (Mask[i]-e == i);
12367 // Eliminate identity shuffles.
12368 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
12369 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
12371 // If the LHS is a shufflevector itself, see if we can combine it with this
12372 // one without producing an unusual shuffle. Here we are really conservative:
12373 // we are absolutely afraid of producing a shuffle mask not in the input
12374 // program, because the code gen may not be smart enough to turn a merged
12375 // shuffle into two specific shuffles: it may produce worse code. As such,
12376 // we only merge two shuffles if the result is one of the two input shuffle
12377 // masks. In this case, merging the shuffles just removes one instruction,
12378 // which we know is safe. This is good for things like turning:
12379 // (splat(splat)) -> splat.
12380 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
12381 if (isa<UndefValue>(RHS)) {
12382 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
12384 std::vector<unsigned> NewMask;
12385 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
12386 if (Mask[i] >= 2*e)
12387 NewMask.push_back(2*e);
12389 NewMask.push_back(LHSMask[Mask[i]]);
12391 // If the result mask is equal to the src shuffle or this shuffle mask, do
12392 // the replacement.
12393 if (NewMask == LHSMask || NewMask == Mask) {
12394 unsigned LHSInNElts =
12395 cast<VectorType>(LHSSVI->getOperand(0)->getType())->getNumElements();
12396 std::vector<Constant*> Elts;
12397 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
12398 if (NewMask[i] >= LHSInNElts*2) {
12399 Elts.push_back(UndefValue::get(Type::Int32Ty));
12401 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
12404 return new ShuffleVectorInst(LHSSVI->getOperand(0),
12405 LHSSVI->getOperand(1),
12406 ConstantVector::get(Elts));
12411 return MadeChange ? &SVI : 0;
12417 /// TryToSinkInstruction - Try to move the specified instruction from its
12418 /// current block into the beginning of DestBlock, which can only happen if it's
12419 /// safe to move the instruction past all of the instructions between it and the
12420 /// end of its block.
12421 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
12422 assert(I->hasOneUse() && "Invariants didn't hold!");
12424 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
12425 if (isa<PHINode>(I) || I->mayWriteToMemory() || isa<TerminatorInst>(I))
12428 // Do not sink alloca instructions out of the entry block.
12429 if (isa<AllocaInst>(I) && I->getParent() ==
12430 &DestBlock->getParent()->getEntryBlock())
12433 // We can only sink load instructions if there is nothing between the load and
12434 // the end of block that could change the value.
12435 if (I->mayReadFromMemory()) {
12436 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
12438 if (Scan->mayWriteToMemory())
12442 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
12444 CopyPrecedingStopPoint(I, InsertPos);
12445 I->moveBefore(InsertPos);
12451 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
12452 /// all reachable code to the worklist.
12454 /// This has a couple of tricks to make the code faster and more powerful. In
12455 /// particular, we constant fold and DCE instructions as we go, to avoid adding
12456 /// them to the worklist (this significantly speeds up instcombine on code where
12457 /// many instructions are dead or constant). Additionally, if we find a branch
12458 /// whose condition is a known constant, we only visit the reachable successors.
12460 static void AddReachableCodeToWorklist(BasicBlock *BB,
12461 SmallPtrSet<BasicBlock*, 64> &Visited,
12463 const TargetData *TD) {
12464 SmallVector<BasicBlock*, 256> Worklist;
12465 Worklist.push_back(BB);
12467 while (!Worklist.empty()) {
12468 BB = Worklist.back();
12469 Worklist.pop_back();
12471 // We have now visited this block! If we've already been here, ignore it.
12472 if (!Visited.insert(BB)) continue;
12474 DbgInfoIntrinsic *DBI_Prev = NULL;
12475 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
12476 Instruction *Inst = BBI++;
12478 // DCE instruction if trivially dead.
12479 if (isInstructionTriviallyDead(Inst)) {
12481 DOUT << "IC: DCE: " << *Inst;
12482 Inst->eraseFromParent();
12486 // ConstantProp instruction if trivially constant.
12487 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
12488 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
12489 Inst->replaceAllUsesWith(C);
12491 Inst->eraseFromParent();
12495 // If there are two consecutive llvm.dbg.stoppoint calls then
12496 // it is likely that the optimizer deleted code in between these
12498 DbgInfoIntrinsic *DBI_Next = dyn_cast<DbgInfoIntrinsic>(Inst);
12501 && DBI_Prev->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint
12502 && DBI_Next->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint) {
12503 IC.RemoveFromWorkList(DBI_Prev);
12504 DBI_Prev->eraseFromParent();
12506 DBI_Prev = DBI_Next;
12511 IC.AddToWorkList(Inst);
12514 // Recursively visit successors. If this is a branch or switch on a
12515 // constant, only visit the reachable successor.
12516 TerminatorInst *TI = BB->getTerminator();
12517 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
12518 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
12519 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
12520 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
12521 Worklist.push_back(ReachableBB);
12524 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
12525 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
12526 // See if this is an explicit destination.
12527 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
12528 if (SI->getCaseValue(i) == Cond) {
12529 BasicBlock *ReachableBB = SI->getSuccessor(i);
12530 Worklist.push_back(ReachableBB);
12534 // Otherwise it is the default destination.
12535 Worklist.push_back(SI->getSuccessor(0));
12540 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
12541 Worklist.push_back(TI->getSuccessor(i));
12545 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
12546 bool Changed = false;
12547 TD = &getAnalysis<TargetData>();
12549 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
12550 << F.getNameStr() << "\n");
12553 // Do a depth-first traversal of the function, populate the worklist with
12554 // the reachable instructions. Ignore blocks that are not reachable. Keep
12555 // track of which blocks we visit.
12556 SmallPtrSet<BasicBlock*, 64> Visited;
12557 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
12559 // Do a quick scan over the function. If we find any blocks that are
12560 // unreachable, remove any instructions inside of them. This prevents
12561 // the instcombine code from having to deal with some bad special cases.
12562 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
12563 if (!Visited.count(BB)) {
12564 Instruction *Term = BB->getTerminator();
12565 while (Term != BB->begin()) { // Remove instrs bottom-up
12566 BasicBlock::iterator I = Term; --I;
12568 DOUT << "IC: DCE: " << *I;
12569 // A debug intrinsic shouldn't force another iteration if we weren't
12570 // going to do one without it.
12571 if (!isa<DbgInfoIntrinsic>(I)) {
12575 if (!I->use_empty())
12576 I->replaceAllUsesWith(UndefValue::get(I->getType()));
12577 I->eraseFromParent();
12582 while (!Worklist.empty()) {
12583 Instruction *I = RemoveOneFromWorkList();
12584 if (I == 0) continue; // skip null values.
12586 // Check to see if we can DCE the instruction.
12587 if (isInstructionTriviallyDead(I)) {
12588 // Add operands to the worklist.
12589 if (I->getNumOperands() < 4)
12590 AddUsesToWorkList(*I);
12593 DOUT << "IC: DCE: " << *I;
12595 I->eraseFromParent();
12596 RemoveFromWorkList(I);
12601 // Instruction isn't dead, see if we can constant propagate it.
12602 if (Constant *C = ConstantFoldInstruction(I, TD)) {
12603 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
12605 // Add operands to the worklist.
12606 AddUsesToWorkList(*I);
12607 ReplaceInstUsesWith(*I, C);
12610 I->eraseFromParent();
12611 RemoveFromWorkList(I);
12616 if (TD && I->getType()->getTypeID() == Type::VoidTyID) {
12617 // See if we can constant fold its operands.
12618 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
12619 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i))
12620 if (Constant *NewC = ConstantFoldConstantExpression(CE, TD))
12627 // See if we can trivially sink this instruction to a successor basic block.
12628 if (I->hasOneUse()) {
12629 BasicBlock *BB = I->getParent();
12630 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
12631 if (UserParent != BB) {
12632 bool UserIsSuccessor = false;
12633 // See if the user is one of our successors.
12634 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
12635 if (*SI == UserParent) {
12636 UserIsSuccessor = true;
12640 // If the user is one of our immediate successors, and if that successor
12641 // only has us as a predecessors (we'd have to split the critical edge
12642 // otherwise), we can keep going.
12643 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
12644 next(pred_begin(UserParent)) == pred_end(UserParent))
12645 // Okay, the CFG is simple enough, try to sink this instruction.
12646 Changed |= TryToSinkInstruction(I, UserParent);
12650 // Now that we have an instruction, try combining it to simplify it...
12654 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
12655 if (Instruction *Result = visit(*I)) {
12657 // Should we replace the old instruction with a new one?
12659 DOUT << "IC: Old = " << *I
12660 << " New = " << *Result;
12662 // Everything uses the new instruction now.
12663 I->replaceAllUsesWith(Result);
12665 // Push the new instruction and any users onto the worklist.
12666 AddToWorkList(Result);
12667 AddUsersToWorkList(*Result);
12669 // Move the name to the new instruction first.
12670 Result->takeName(I);
12672 // Insert the new instruction into the basic block...
12673 BasicBlock *InstParent = I->getParent();
12674 BasicBlock::iterator InsertPos = I;
12676 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
12677 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
12680 InstParent->getInstList().insert(InsertPos, Result);
12682 // Make sure that we reprocess all operands now that we reduced their
12684 AddUsesToWorkList(*I);
12686 // Instructions can end up on the worklist more than once. Make sure
12687 // we do not process an instruction that has been deleted.
12688 RemoveFromWorkList(I);
12690 // Erase the old instruction.
12691 InstParent->getInstList().erase(I);
12694 DOUT << "IC: Mod = " << OrigI
12695 << " New = " << *I;
12698 // If the instruction was modified, it's possible that it is now dead.
12699 // if so, remove it.
12700 if (isInstructionTriviallyDead(I)) {
12701 // Make sure we process all operands now that we are reducing their
12703 AddUsesToWorkList(*I);
12705 // Instructions may end up in the worklist more than once. Erase all
12706 // occurrences of this instruction.
12707 RemoveFromWorkList(I);
12708 I->eraseFromParent();
12711 AddUsersToWorkList(*I);
12718 assert(WorklistMap.empty() && "Worklist empty, but map not?");
12720 // Do an explicit clear, this shrinks the map if needed.
12721 WorklistMap.clear();
12726 bool InstCombiner::runOnFunction(Function &F) {
12727 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
12729 bool EverMadeChange = false;
12731 // Iterate while there is work to do.
12732 unsigned Iteration = 0;
12733 while (DoOneIteration(F, Iteration++))
12734 EverMadeChange = true;
12735 return EverMadeChange;
12738 FunctionPass *llvm::createInstructionCombiningPass() {
12739 return new InstCombiner();